From d202707ff23e677dbed0272047ae9ff938652592 Mon Sep 17 00:00:00 2001 From: Alexis Rolland Date: Wed, 17 Jun 2026 21:02:45 +0800 Subject: [PATCH 01/65] Update TripoSplat categories (#14512) --- comfy_extras/nodes_triposplat.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/comfy_extras/nodes_triposplat.py b/comfy_extras/nodes_triposplat.py index 1848ad31a..7bf4703fe 100644 --- a/comfy_extras/nodes_triposplat.py +++ b/comfy_extras/nodes_triposplat.py @@ -65,7 +65,7 @@ class TripoSplatPreprocessImage(IO.ComfyNode): return IO.Schema( node_id="TripoSplatPreprocessImage", display_name="TripoSplat Preprocess Image", - category="3d/conditioning", + category="model/conditioning/triposplat", description="Crop center each image to a square canvas on a black background and add padding.", inputs=[ IO.Image.Input("image"), @@ -95,7 +95,7 @@ class TripoSplatConditioning(IO.ComfyNode): return IO.Schema( node_id="TripoSplatConditioning", display_name="TripoSplat Conditioning", - category="3d/conditioning", + category="model/conditioning/triposplat", description="Encode the image with DINOv3 and the Flux2 VAE into TripoSplat positive/negative " "conditioning, and create the fixed size noise target (latent + camera) for the KSampler", inputs=[ From a590d60bb1d7d47c1cdb49fc8116b0e919fc4bd1 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Jukka=20Sepp=C3=A4nen?= <40791699+kijai@users.noreply.github.com> Date: Wed, 17 Jun 2026 16:21:23 +0300 Subject: [PATCH 02/65] feat: SCAIL-2 multireference (CORE-310) (#14509) * SCAIl-2: support multiref --- comfy/ldm/wan/model.py | 19 +++---- comfy/model_base.py | 20 ++++--- comfy_extras/nodes_scail.py | 110 ++++++++++++++++++++++-------------- 3 files changed, 91 insertions(+), 58 deletions(-) diff --git a/comfy/ldm/wan/model.py b/comfy/ldm/wan/model.py index 282408891..1c9782a38 100644 --- a/comfy/ldm/wan/model.py +++ b/comfy/ldm/wan/model.py @@ -1665,7 +1665,7 @@ class SCAILWanModel(WanModel): # embeddings x = self.patch_embedding(x.float()).to(x.dtype) - if ref_mask_latents is not None: # SCAIL-2 additive mask stream + if ref_mask_latents is not None: # SCAIL-2 additive mask stream (one identity mask frame per reference, then video) x = x + self.patch_embedding_mask(ref_mask_latents.float()).to(x.dtype) grid_sizes = x.shape[2:] transformer_options["grid_sizes"] = grid_sizes @@ -1728,22 +1728,25 @@ class SCAILWanModel(WanModel): # ref_mask_flag is a scalar bool (CONDConstant, SCAIL-2 only). False => replacement mode, # which places ref/pose via H/W rope shifts instead of the animation-mode temporal offset. + # reference_latent may stack several frames: the last is the primary reference adjacent to the video, the earlier frames are additional references. def rope_encode(self, t, h, w, t_start=0, steps_t=None, steps_h=None, steps_w=None, device=None, dtype=None, pose_latents=None, reference_latent=None, ref_mask_flag=None, transformer_options={}): + ref_t_patches = 0 + if reference_latent is not None: + ref_t_patches = (reference_latent.shape[2] + (self.patch_size[0] // 2)) // self.patch_size[0] + if ref_mask_flag is not None and not bool(ref_mask_flag): REF_ROPE_H = 120.0 POSE_ROPE_W = 120.0 - ref_t_patches = 0 - if reference_latent is not None: - ref_t_patches = (reference_latent.shape[2] + (self.patch_size[0] // 2)) // self.patch_size[0] main_t_patches = t - ref_t_patches + video_t_start = max(ref_t_patches - 1, 0) parts = [] if ref_t_patches > 0: ref_tf = {"rope_options": {"shift_y": REF_ROPE_H, "shift_x": 0.0, "scale_y": 1.0, "scale_x": 1.0}} parts.append(super().rope_encode(ref_t_patches, h, w, t_start=0, device=device, dtype=dtype, transformer_options=ref_tf)) if main_t_patches > 0: - parts.append(super().rope_encode(main_t_patches, h, w, t_start=0, device=device, dtype=dtype, transformer_options=transformer_options)) + parts.append(super().rope_encode(main_t_patches, h, w, t_start=video_t_start, device=device, dtype=dtype, transformer_options=transformer_options)) if pose_latents is not None: F_pose, H_pose, W_pose = pose_latents.shape[-3], pose_latents.shape[-2], pose_latents.shape[-1] @@ -1752,7 +1755,7 @@ class SCAILWanModel(WanModel): h_shift = (h_scale - 1) / 2 w_shift = (w_scale - 1) / 2 pose_tf = {"rope_options": {"shift_y": h_shift, "shift_x": POSE_ROPE_W + w_shift, "scale_y": h_scale, "scale_x": w_scale}} - parts.append(super().rope_encode(F_pose, H_pose, W_pose, t_start=0, device=device, dtype=dtype, transformer_options=pose_tf)) + parts.append(super().rope_encode(F_pose, H_pose, W_pose, t_start=video_t_start, device=device, dtype=dtype, transformer_options=pose_tf)) return torch.cat(parts, dim=1) @@ -1761,10 +1764,6 @@ class SCAILWanModel(WanModel): if pose_latents is None: return main_freqs - ref_t_patches = 0 - if reference_latent is not None: - ref_t_patches = (reference_latent.shape[2] + (self.patch_size[0] // 2)) // self.patch_size[0] - F_pose, H_pose, W_pose = pose_latents.shape[-3], pose_latents.shape[-2], pose_latents.shape[-1] # if pose is at half resolution, scale_y/scale_x=2 stretches the position range to cover the same RoPE extent as the main frames diff --git a/comfy/model_base.py b/comfy/model_base.py index ab4a11022..d143dc06f 100644 --- a/comfy/model_base.py +++ b/comfy/model_base.py @@ -1747,10 +1747,14 @@ class WAN21_SCAIL(WAN21): reference_latents = kwargs.get("reference_latents", None) if reference_latents is not None: - ref_latent = self.process_latent_in(reference_latents[-1]) - ref_mask = torch.ones_like(ref_latent[:, :4]) - ref_latent = torch.cat([ref_latent, ref_mask], dim=1) - out['reference_latent'] = comfy.conds.CONDRegular(ref_latent) + # SCAIL-2 multi-reference: reference_latents[0] is the primary ref, [1:] are additional + # references. Stack as [additional..., primary] so the primary stays adjacent to the video. + ordered = list(reference_latents[1:]) + list(reference_latents[:1]) + stacked = [] + for lat in ordered: + lat = self.process_latent_in(lat) + stacked.append(torch.cat([lat, torch.ones_like(lat[:, :4])], dim=1)) + out['reference_latent'] = comfy.conds.CONDRegular(torch.cat(stacked, dim=2)) pose_latents = kwargs.get("pose_video_latent", None) if pose_latents is not None: @@ -1792,6 +1796,7 @@ class WAN21_SCAIL2(WAN21_SCAIL): if driving_mask_28ch is not None: out['sam_latents'] = comfy.conds.CONDRegular(driving_mask_28ch.movedim(1, 2).contiguous()) + # ref_mask_28ch holds one identity mask per stacked reference frame (additional refs first, then the primary ref), followed by zeros over the video frames. ref_mask_28ch = kwargs.get("ref_mask_28ch", None) if ref_mask_28ch is not None: out['ref_mask_latents'] = comfy.conds.CONDRegular(ref_mask_28ch.movedim(1, 2).contiguous()) @@ -1819,10 +1824,11 @@ class WAN21_SCAIL2(WAN21_SCAIL): # Return sliced view omitting retain_index_list return comfy.context_windows.slice_cond(cond_value, window, x_in, device, temporal_dim=2, temporal_offset=0) if cond_key == "ref_mask_latents" and hasattr(cond_value, "cond") and isinstance(cond_value.cond, torch.Tensor): - # The ref mask is just a single frame padded with frames of zeros, so just grab the first frames for all windows + # The ref mask is N leading ref frames padded with frames of zeros, so just grab the first frames for all windows full_ref_mask = cond_value.cond video_frame_count = x_in.shape[2] - if full_ref_mask.shape[2] != video_frame_count + 1: + ref_frame_count = full_ref_mask.shape[2] - video_frame_count + if ref_frame_count < 1: return None window_length = len(window.index_list) @@ -1831,7 +1837,7 @@ class WAN21_SCAIL2(WAN21_SCAIL): if anchor_index is not None and anchor_index >= 0: window_length += 1 - window_ref_mask = full_ref_mask[:, :, :window_length + 1].to(device) + window_ref_mask = full_ref_mask[:, :, :window_length + ref_frame_count].to(device) return cond_value._copy_with(window_ref_mask) return super().resize_cond_for_context_window(cond_key, cond_value, window, x_in, device, retain_index_list=retain_index_list) diff --git a/comfy_extras/nodes_scail.py b/comfy_extras/nodes_scail.py index 007733efc..55c9897e3 100644 --- a/comfy_extras/nodes_scail.py +++ b/comfy_extras/nodes_scail.py @@ -34,14 +34,20 @@ def _unpack(track_data): return unpack_masks(packed) -def _first_frame_cx_area(masks_bool): - first = masks_bool[0].float() - H, W = first.shape[-2], first.shape[-1] - n_pixels = H * W - grid_x = torch.arange(W, device=first.device, dtype=first.dtype).view(1, W) - area = first.sum(dim=(-1, -2)).clamp_(min=1) - cx = (first * grid_x).sum(dim=(-1, -2)) / area - return (cx / W).tolist(), (area / n_pixels).tolist() +def _first_appearance_cx_area(masks_bool): + """Per object: first frame it appears in, plus centroid-x and area in that frame.""" + m = masks_bool.float() + T, H, W = m.shape[0], m.shape[-2], m.shape[-1] + grid_x = torch.arange(W, device=m.device, dtype=m.dtype).view(1, 1, 1, W) + area_t = m.sum(dim=(-1, -2)) + cx_t = (m * grid_x).sum(dim=(-1, -2)) / area_t.clamp(min=1) + present = area_t > 0 + frame_idx = torch.arange(T, device=m.device).unsqueeze(1) + first_t = torch.where(present, frame_idx, T).amin(dim=0) + sel = first_t.clamp(max=T - 1).unsqueeze(0) + cx = cx_t.gather(0, sel).squeeze(0) + area = area_t.gather(0, sel).squeeze(0) + return first_t.tolist(), (cx / W).tolist(), (area / (H * W)).tolist() def _subset_track_data(track_data, obj_indices): @@ -81,12 +87,26 @@ def _render_colored_masks(track_data, background="black"): masks_full.view(T * N_obj, 1, Hm, Wm), size=(H, W), mode="nearest" ).view(T, N_obj, H, W) > 0.5 any_mask = masks_full.any(dim=1) - obj_idx_map = masks_full.to(torch.uint8).argmax(dim=1) - color_overlay = colors[obj_idx_map] + color_overlay = colors[masks_full.to(torch.uint8).argmax(dim=1)] bg_tensor = torch.tensor(bg_rgb, device=device, dtype=color_overlay.dtype).view(1, 1, 1, 3) return torch.where(any_mask.unsqueeze(-1), color_overlay, bg_tensor.expand_as(color_overlay)) +def _render_mask_as_identity(mask, background="black"): + """Plain comfy MASK (B,H,W) or (H,W) -> (B,H,W,3) rendered as a single identity (palette[0]) + on the given background. A batch is treated as multiple views of that one subject.""" + device = comfy.model_management.intermediate_device() + dtype = comfy.model_management.intermediate_dtype() + if mask.ndim == 2: + mask = mask.unsqueeze(0) + mask = mask.to(device=device, dtype=dtype) + B, H, W = mask.shape + bg_rgb = (1.0, 1.0, 1.0) if background.startswith("white") else (0.0, 0.0, 0.0) + color = torch.tensor(DEFAULT_PALETTE[0], device=device, dtype=dtype).view(1, 1, 1, 3) + bg = torch.tensor(bg_rgb, device=device, dtype=dtype).view(1, 1, 1, 3) + return torch.where((mask > 0.5).unsqueeze(-1), color.expand(B, H, W, 3), bg.expand(B, H, W, 3)) + + def _extract_mask_to_28ch(rgb_video): """Colored RGB mask (T, H, W, 3) in [0, 1] -> SCAIL-2 28-channel binary latent (1, T_lat, 28, H_lat, W_lat). 7 per-color binary channels (white/r/g/b/y/m/c) @@ -138,8 +158,8 @@ class WanSCAILToVideo(io.ComfyNode): io.Float.Input("pose_strength", default=1.0, min=0.0, max=10.0, step=0.01, tooltip="Strength of the pose latent."), io.Float.Input("pose_start", default=0.0, min=0.0, max=1.0, step=0.01, tooltip="Start step of the pose conditioning."), io.Float.Input("pose_end", default=1.0, min=0.0, max=1.0, step=0.01, tooltip="End step of the pose conditioning."), - io.Image.Input("reference_image", optional=True, tooltip="Reference image, for multiple references composite all on single image."), - io.Image.Input("reference_image_mask", optional=True, tooltip="SCAIL-2 only. Colored reference mask at the same resolution as reference_image."), + io.Image.Input("reference_image", optional=True, tooltip="Reference image. The first image is the primary reference (composite all identities onto it). SCAIL-2: extra batch images are used as additional views (back view, close-up, occluded background), each needing a matching reference_image_mask in that identity's color."), + io.Image.Input("reference_image_mask", optional=True, tooltip="SCAIL-2 only. Colored reference mask, batch matching reference_image (first = primary reference mask, rest = identity masks for the additional reference_image)."), io.ClipVisionOutput.Input("clip_vision_output", optional=True, tooltip="CLIP vision features for conditioning. Model is trained with stretch resize to aspect ratio."), io.Int.Input("video_frame_offset", default=0, min=0, max=nodes.MAX_RESOLUTION, step=1, tooltip="Cumulative output frame this chunk begins at. Wire from the previous chunk's video_frame_offset output."), io.Int.Input("previous_frame_count", default=5, min=1, max=nodes.MAX_RESOLUTION, step=4, tooltip="Tail frames of previous_frames to anchor. SCAIL-2 trained at 5 (81-frame chunks, 76-frame step)."), @@ -171,19 +191,21 @@ class WanSCAILToVideo(io.ComfyNode): video_frame_offset -= prev_trimmed.shape[0] video_frame_offset = max(0, video_frame_offset) - ref_latent = None if reference_image is not None: - reference_image = comfy.utils.common_upscale(reference_image[:1].movedim(-1, 1), width, height, "bicubic", "center").movedim(1, -1) - # Replacement Mode: composite ref on black bg using reference_image_mask as alpha matte - if replacement_mode and reference_image_mask is not None: - rm = comfy.utils.common_upscale(reference_image_mask[:1].movedim(-1, 1), width, height, "nearest-exact", "center").movedim(1, -1) - is_char = (rm[..., :3].max(dim=-1, keepdim=True).values > 0.1).to(reference_image.dtype) - reference_image = reference_image * is_char - ref_latent = vae.encode(reference_image[:, :, :, :3]) + ref_imgs = comfy.utils.common_upscale(reference_image.movedim(-1, 1), width, height, "bicubic", "center").movedim(1, -1) + n_ref = ref_imgs.shape[0] + # SCAIL-2 multi-reference: the first image is the primary ref, the rest are additional references. - if ref_latent is not None: - positive = node_helpers.conditioning_set_values(positive, {"reference_latents": [ref_latent]}, append=True) - negative = node_helpers.conditioning_set_values(negative, {"reference_latents": [ref_latent]}, append=True) + # Replacement Mode: composite each ref on black bg using its mask as alpha matte + if replacement_mode and reference_image_mask is not None: + rm = comfy.utils.common_upscale(reference_image_mask.movedim(-1, 1), width, height, "nearest-exact", "center").movedim(1, -1) + rm = rm[[min(i, rm.shape[0] - 1) for i in range(n_ref)]] + is_char = (rm[..., :3].max(dim=-1, keepdim=True).values > 0.1).to(ref_imgs.dtype) + ref_imgs = ref_imgs * is_char + # encode each ref individually so each stays a single latent frame (a batched encode would be treated as a video) + ref_latents = [vae.encode(ref_imgs[i:i + 1, :, :, :3]) for i in range(n_ref)] + positive = node_helpers.conditioning_set_values(positive, {"reference_latents": ref_latents}, append=True) + negative = node_helpers.conditioning_set_values(negative, {"reference_latents": ref_latents}, append=True) if clip_vision_output is not None: positive = node_helpers.conditioning_set_values(positive, {"clip_vision_output": clip_vision_output}) @@ -221,11 +243,16 @@ class WanSCAILToVideo(io.ComfyNode): positive = node_helpers.conditioning_set_values(positive, {"driving_mask_28ch": driving_mask_28ch}) negative = node_helpers.conditioning_set_values(negative, {"driving_mask_28ch": driving_mask_28ch}) - if reference_image_mask is not None: - ref_mask_hw = comfy.utils.common_upscale(reference_image_mask[:1].movedim(-1, 1), width, height, "bicubic", "center").movedim(1, -1) - ref_mask_1f = _extract_mask_to_28ch(ref_mask_hw) + # The ref mask binds reference frames to identities, so it only applies when there's a reference image. + if reference_image_mask is not None and reference_image is not None: + ref_mask_hw = comfy.utils.common_upscale(reference_image_mask.movedim(-1, 1), width, height, "nearest-exact", "center").movedim(1, -1) + n_masks = ref_mask_hw.shape[0] + n_ref = reference_image.shape[0] + + add_masks = [_extract_mask_to_28ch(ref_mask_hw[min(i, n_masks - 1)][None]) for i in range(1, n_ref)] + ref_mask_1f = _extract_mask_to_28ch(ref_mask_hw[:1]) zeros = torch.zeros((1, latent.shape[2], 28, ref_mask_1f.shape[-2], ref_mask_1f.shape[-1]), device=ref_mask_1f.device, dtype=ref_mask_1f.dtype) - ref_mask_28ch = torch.cat([ref_mask_1f, zeros], dim=1) + ref_mask_28ch = torch.cat(add_masks + [ref_mask_1f, zeros], dim=1) positive = node_helpers.conditioning_set_values(positive, {"ref_mask_28ch": ref_mask_28ch}) negative = node_helpers.conditioning_set_values(negative, {"ref_mask_28ch": ref_mask_28ch}) @@ -244,12 +271,9 @@ class WanSCAILToVideo(io.ComfyNode): class SCAIL2ColoredMask(io.ComfyNode): - """Render SAM3 tracks for the driving pose video and (optionally) the reference - image into the two colored masks WanSCAILToVideo consumes. Shared `sort_by` - across both outputs guarantees identity K maps to the same color on both - sides, for multi-person workflow consistency. - reference_image_mask is always rendered black-bg (model convention) - pose_video_mask bg follows replacement_mode: black = Animation Mode, white = Replacement Mode + """Render SAM3 tracks for the driving pose video and reference image(s) into the + colored masks WanSCAILToVideo consumes. Shared `sort_by` keeps each identity on the + same color across both outputs. """ @classmethod @@ -260,10 +284,12 @@ class SCAIL2ColoredMask(io.ComfyNode): category="model/conditioning/wan/scail", inputs=[ SAM3TrackData.Input("driving_track_data", tooltip="SAM3 track of the driving pose video. Will be rendered into the pose_video_mask output."), - SAM3TrackData.Input("ref_track_data", optional=True, tooltip="SAM3 track of the reference image."), - io.String.Input("object_indices", default="", tooltip="Comma-separated list of person indices to include (e.g. '0,2,3'). Applied to both reference and pose video masks. Empty = all."), + io.MultiType.Input("ref_track_data", [SAM3TrackData, io.Mask], optional=True, display_name="reference_masks", + tooltip="SAM3 track of the reference image(s) (one identity per object, colored in batch order), or a plain MASK of the reference subject (rendered as a single identity)."), + io.String.Input("object_indices", default="", + tooltip="Comma-separated list of person indices to include (e.g. '0,2,3'). Applied to both reference and pose video masks. Empty = all."), io.Combo.Input("sort_by", options=["none", "left_to_right", "area"], default="left_to_right", - tooltip="Order in which palette colors are assigned to the tracked objects (applied to both reference and pose video so each identity keeps the same color). left_to_right = leftmost object (by first-frame centroid) gets the first color; area = biggest object (by first-frame mask area) gets the first color; none = keep SAM3's order."), + tooltip="Order in which palette colors are assigned to the tracked objects (applied to both reference and pose video so each identity keeps the same color). Objects that appear in earlier frames always come first; within a frame, left_to_right = leftmost object (by centroid at first appearance) gets the first color, area = biggest object (by mask area at first appearance) gets the first color; none = keep SAM3's order."), io.Boolean.Input("replacement_mode", default=False, tooltip="False = Animation Mode (pose_video_mask has black background, reference_image_mask has white background). " "True = Replacement Mode (pose_video_mask has white background, reference_image_mask has black background)."), @@ -280,11 +306,11 @@ class SCAIL2ColoredMask(io.ComfyNode): def _prep(td): masks_bool = _unpack(td) if sort_by != "none" and masks_bool is not None: - cx, area = _first_frame_cx_area(masks_bool) + first_t, cx, area = _first_appearance_cx_area(masks_bool) if sort_by == "left_to_right": - order = sorted(range(len(cx)), key=lambda i: cx[i]) + order = sorted(range(len(cx)), key=lambda i: (first_t[i], cx[i])) else: # "area" - order = sorted(range(len(area)), key=lambda i: -area[i]) + order = sorted(range(len(area)), key=lambda i: (first_t[i], -area[i])) td = _subset_track_data(td, order) if object_indices.strip(): indices = [int(i.strip()) for i in object_indices.split(",") if i.strip().isdigit()] @@ -300,8 +326,10 @@ class SCAIL2ColoredMask(io.ComfyNode): ref_bg = "black" if replacement_mode else "white" if ref_track_data is not None: - ref = _prep(ref_track_data) - reference_image_mask = _render_colored_masks(ref, ref_bg) + if isinstance(ref_track_data, torch.Tensor): # plain comfy MASK + reference_image_mask = _render_mask_as_identity(ref_track_data, ref_bg) + else: + reference_image_mask = _render_colored_masks(_prep(ref_track_data), ref_bg) else: H, W = drv["orig_size"] fill_value = 1.0 if ref_bg == "white" else 0.0 From ca3dbe206c2fea84f2af4371ca13e9f2bfeb82e5 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Jukka=20Sepp=C3=A4nen?= <40791699+kijai@users.noreply.github.com> Date: Wed, 17 Jun 2026 18:45:06 +0300 Subject: [PATCH 03/65] Allow using Qwen3-VL as flux2 klein text encoder (again) (#14526) --- comfy/sd.py | 4 ++++ 1 file changed, 4 insertions(+) diff --git a/comfy/sd.py b/comfy/sd.py index 688e6db90..348fe4958 100644 --- a/comfy/sd.py +++ b/comfy/sd.py @@ -1622,6 +1622,10 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip clip_data[0] = comfy.utils.state_dict_prefix_replace(clip_data[0], {"model.language_model.": "model.", "model.visual.": "visual.", "lm_head.": "model.lm_head."}) clip_target.clip = comfy.text_encoders.ideogram4.te_qwen3vl(**llama_detect(clip_data)) clip_target.tokenizer = comfy.text_encoders.ideogram4.Ideogram4Qwen3VLTokenizer + elif clip_type in (CLIPType.FLUX, CLIPType.FLUX2): # Flux2 Klein reuses the Qwen3-VL LM (3-layer tap -> 12288); visual unused. + klein_model_type = "qwen3_8b" if te_model == TEModel.QWEN3VL_8B else "qwen3_4b" + clip_target.clip = comfy.text_encoders.flux.klein_te(**llama_detect(clip_data), model_type=klein_model_type) + clip_target.tokenizer = comfy.text_encoders.flux.KleinTokenizer8B if te_model == TEModel.QWEN3VL_8B else comfy.text_encoders.flux.KleinTokenizer else: clip_data[0] = comfy.utils.state_dict_prefix_replace(clip_data[0], {"model.language_model.": "model.", "model.visual.": "visual.", "lm_head.": "model.lm_head."}) qwen3vl_type = {TEModel.QWEN3VL_4B: "qwen3vl_4b", TEModel.QWEN3VL_8B: "qwen3vl_8b"}[te_model] From e25c391888618967f7ef199f466ff29da2ed14b3 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Jukka=20Sepp=C3=A4nen?= <40791699+kijai@users.noreply.github.com> Date: Thu, 18 Jun 2026 00:22:36 +0300 Subject: [PATCH 04/65] feat: Support Boogu-Image (CORE-308) (#14523) --- comfy/ldm/boogu/model.py | 321 ++++++++++++++++++++++++++++++++++ comfy/ldm/omnigen/omnigen2.py | 2 +- comfy/model_base.py | 6 + comfy/model_detection.py | 10 ++ comfy/sd.py | 6 + comfy/supported_models.py | 23 +++ comfy/text_encoders/boogu.py | 58 ++++++ comfy_extras/nodes_boogu.py | 96 ++++++++++ nodes.py | 3 +- 9 files changed, 523 insertions(+), 2 deletions(-) create mode 100644 comfy/ldm/boogu/model.py create mode 100644 comfy/text_encoders/boogu.py create mode 100644 comfy_extras/nodes_boogu.py diff --git a/comfy/ldm/boogu/model.py b/comfy/ldm/boogu/model.py new file mode 100644 index 000000000..966f3c583 --- /dev/null +++ b/comfy/ldm/boogu/model.py @@ -0,0 +1,321 @@ +# Boogu-Image-0.1 transformer +# Architecture is an OmniGen2 derivative (see comfy/ldm/omnigen/omnigen2.py) with an +# added dual-stream ("double_stream") stage before the single-stream layers, conditioned +# by a Qwen3-VL multimodal LLM. Reuses the OmniGen2/Lumina building blocks and the Flux +# RoPE core, the only new component is the double-stream block + the hybrid forward order. + +from typing import Optional, Tuple + +import torch +import torch.nn as nn +from einops import rearrange + +import comfy.ldm.common_dit +import comfy.ldm.omnigen.omnigen2 +from comfy.ldm.modules.attention import optimized_attention_masked +from comfy.ldm.omnigen.omnigen2 import ( + OmniGen2RotaryPosEmbed, + Lumina2CombinedTimestepCaptionEmbedding, + LuminaRMSNormZero, + LuminaLayerNormContinuous, + LuminaFeedForward, + Attention, + OmniGen2TransformerBlock, + apply_rotary_emb, +) + +class BooguDoubleStreamProcessor(nn.Module): + # Joint attention over [instruct ; img] with separate per-stream q/k/v and output projections. + def __init__(self, dim, head_dim, heads, kv_heads, dtype=None, device=None, operations=None): + super().__init__() + query_dim = head_dim * heads + kv_dim = head_dim * kv_heads + + self.img_to_q = operations.Linear(query_dim, query_dim, bias=False, dtype=dtype, device=device) + self.img_to_k = operations.Linear(query_dim, kv_dim, bias=False, dtype=dtype, device=device) + self.img_to_v = operations.Linear(query_dim, kv_dim, bias=False, dtype=dtype, device=device) + + self.instruct_to_q = operations.Linear(query_dim, query_dim, bias=False, dtype=dtype, device=device) + self.instruct_to_k = operations.Linear(query_dim, kv_dim, bias=False, dtype=dtype, device=device) + self.instruct_to_v = operations.Linear(query_dim, kv_dim, bias=False, dtype=dtype, device=device) + + self.instruct_out = operations.Linear(query_dim, query_dim, bias=False, dtype=dtype, device=device) + self.img_out = operations.Linear(query_dim, query_dim, bias=False, dtype=dtype, device=device) + + def forward(self, attn, img_hidden_states, instruct_hidden_states, rotary_emb, attention_mask=None, transformer_options={}): + batch_size = img_hidden_states.shape[0] + L_instruct = instruct_hidden_states.shape[1] + + img_q = self.img_to_q(img_hidden_states) + img_k = self.img_to_k(img_hidden_states) + img_v = self.img_to_v(img_hidden_states) + + instruct_q = self.instruct_to_q(instruct_hidden_states) + instruct_k = self.instruct_to_k(instruct_hidden_states) + instruct_v = self.instruct_to_v(instruct_hidden_states) + + # Concatenate instruction first, then image (matches reference processor order). + query = torch.cat([instruct_q, img_q], dim=1) + key = torch.cat([instruct_k, img_k], dim=1) + value = torch.cat([instruct_v, img_v], dim=1) + + query = query.view(batch_size, -1, attn.heads, attn.dim_head) + key = key.view(batch_size, -1, attn.kv_heads, attn.dim_head) + value = value.view(batch_size, -1, attn.kv_heads, attn.dim_head) + + query = attn.norm_q(query) + key = attn.norm_k(key) + + if rotary_emb is not None: + query = apply_rotary_emb(query, rotary_emb) + key = apply_rotary_emb(key, rotary_emb) + + query = query.transpose(1, 2) + key = key.transpose(1, 2) + value = value.transpose(1, 2) + + if attn.kv_heads < attn.heads: + key = key.repeat_interleave(attn.heads // attn.kv_heads, dim=1) + value = value.repeat_interleave(attn.heads // attn.kv_heads, dim=1) + + hidden_states = optimized_attention_masked(query, key, value, attn.heads, attention_mask, skip_reshape=True, transformer_options=transformer_options) + + # Split back to instruction/image, apply per-stream output projections, recombine. + instruct_hidden_states = self.instruct_out(hidden_states[:, :L_instruct]) + img_hidden_states = self.img_out(hidden_states[:, L_instruct:]) + hidden_states = torch.cat([instruct_hidden_states, img_hidden_states], dim=1) + + hidden_states = attn.to_out[0](hidden_states) + return hidden_states + + +class BooguJointAttention(nn.Module): + # Holds the shared q/k RMSNorm + final output projection + def __init__(self, dim, head_dim, heads, kv_heads, eps=1e-5, dtype=None, device=None, operations=None): + super().__init__() + self.heads = heads + self.kv_heads = kv_heads + self.dim_head = head_dim + self.scale = head_dim ** -0.5 + + self.norm_q = operations.RMSNorm(head_dim, eps=eps, dtype=dtype, device=device) + self.norm_k = operations.RMSNorm(head_dim, eps=eps, dtype=dtype, device=device) + self.to_out = nn.Sequential( + operations.Linear(heads * head_dim, dim, bias=False, dtype=dtype, device=device), + nn.Dropout(0.0), + ) + self.processor = BooguDoubleStreamProcessor(dim, head_dim, heads, kv_heads, dtype=dtype, device=device, operations=operations) + + def forward(self, img_hidden_states, instruct_hidden_states, rotary_emb, attention_mask=None, transformer_options={}): + return self.processor(self, img_hidden_states, instruct_hidden_states, rotary_emb, attention_mask, transformer_options=transformer_options) + + +class BooguDoubleStreamBlock(nn.Module): + # Dual-stream block: joint attention over [instruct ; img] + image self-attention, each stream with its own modulation/MLP. + def __init__(self, dim, num_attention_heads, num_kv_heads, multiple_of, ffn_dim_multiplier, norm_eps, dtype=None, device=None, operations=None): + super().__init__() + head_dim = dim // num_attention_heads + + self.img_instruct_attn = BooguJointAttention(dim, head_dim, num_attention_heads, num_kv_heads, eps=1e-5, dtype=dtype, device=device, operations=operations) + self.img_self_attn = Attention( + query_dim=dim, dim_head=head_dim, heads=num_attention_heads, kv_heads=num_kv_heads, + eps=1e-5, bias=False, dtype=dtype, device=device, operations=operations, + ) + + self.img_feed_forward = LuminaFeedForward(dim=dim, inner_dim=4 * dim, multiple_of=multiple_of, dtype=dtype, device=device, operations=operations) + self.instruct_feed_forward = LuminaFeedForward(dim=dim, inner_dim=4 * dim, multiple_of=multiple_of, dtype=dtype, device=device, operations=operations) + + self.img_norm1 = LuminaRMSNormZero(embedding_dim=dim, norm_eps=norm_eps, dtype=dtype, device=device, operations=operations) + self.img_norm2 = LuminaRMSNormZero(embedding_dim=dim, norm_eps=norm_eps, dtype=dtype, device=device, operations=operations) + self.img_norm3 = LuminaRMSNormZero(embedding_dim=dim, norm_eps=norm_eps, dtype=dtype, device=device, operations=operations) + self.instruct_norm1 = LuminaRMSNormZero(embedding_dim=dim, norm_eps=norm_eps, dtype=dtype, device=device, operations=operations) + self.instruct_norm2 = LuminaRMSNormZero(embedding_dim=dim, norm_eps=norm_eps, dtype=dtype, device=device, operations=operations) + + self.img_attn_norm = operations.RMSNorm(dim, eps=norm_eps, dtype=dtype, device=device) + self.img_self_attn_norm = operations.RMSNorm(dim, eps=norm_eps, dtype=dtype, device=device) + self.img_ffn_norm1 = operations.RMSNorm(dim, eps=norm_eps, dtype=dtype, device=device) + self.img_ffn_norm2 = operations.RMSNorm(dim, eps=norm_eps, dtype=dtype, device=device) + + self.instruct_attn_norm = operations.RMSNorm(dim, eps=norm_eps, dtype=dtype, device=device) + self.instruct_ffn_norm1 = operations.RMSNorm(dim, eps=norm_eps, dtype=dtype, device=device) + self.instruct_ffn_norm2 = operations.RMSNorm(dim, eps=norm_eps, dtype=dtype, device=device) + + def forward(self, img_hidden_states, instruct_hidden_states, joint_rotary_emb, img_rotary_emb, temb, joint_attention_mask=None, img_attention_mask=None, transformer_options={}): + L_instruct = instruct_hidden_states.shape[1] + + img_norm1_out, img_gate_msa, img_scale_mlp, img_gate_mlp = self.img_norm1(img_hidden_states, temb) + img_norm2_out, img_shift_mlp, _, _ = self.img_norm2(img_hidden_states, temb) + img_norm3_out, img_gate_self, _, _ = self.img_norm3(img_hidden_states, temb) + + instruct_norm1_out, instruct_gate_msa, instruct_scale_mlp, instruct_gate_mlp = self.instruct_norm1(instruct_hidden_states, temb) + instruct_norm2_out, instruct_shift_mlp, _, _ = self.instruct_norm2(instruct_hidden_states, temb) + + joint_attn_out = self.img_instruct_attn(img_norm1_out, instruct_norm1_out, joint_rotary_emb, joint_attention_mask, transformer_options=transformer_options) + instruct_attn_out = joint_attn_out[:, :L_instruct] + img_attn_out = joint_attn_out[:, L_instruct:] + + img_self_attn_out = self.img_self_attn(img_norm3_out, img_norm3_out, img_attention_mask, img_rotary_emb, transformer_options=transformer_options) + + img_hidden_states = img_hidden_states + img_gate_msa.unsqueeze(1).tanh() * self.img_attn_norm(img_attn_out) + img_hidden_states = img_hidden_states + img_gate_self.unsqueeze(1).tanh() * self.img_self_attn_norm(img_self_attn_out) + img_mlp_input = (1 + img_scale_mlp.unsqueeze(1)) * img_norm2_out + img_shift_mlp.unsqueeze(1) + img_mlp_out = self.img_feed_forward(self.img_ffn_norm1(img_mlp_input)) + img_hidden_states = img_hidden_states + img_gate_mlp.unsqueeze(1).tanh() * self.img_ffn_norm2(img_mlp_out) + + instruct_hidden_states = instruct_hidden_states + instruct_gate_msa.unsqueeze(1).tanh() * self.instruct_attn_norm(instruct_attn_out) + instruct_mlp_input = (1 + instruct_scale_mlp.unsqueeze(1)) * instruct_norm2_out + instruct_shift_mlp.unsqueeze(1) + instruct_mlp_out = self.instruct_feed_forward(self.instruct_ffn_norm1(instruct_mlp_input)) + instruct_hidden_states = instruct_hidden_states + instruct_gate_mlp.unsqueeze(1).tanh() * self.instruct_ffn_norm2(instruct_mlp_out) + + return img_hidden_states, instruct_hidden_states + + +class BooguTransformer2DModel(nn.Module): + def __init__( + self, + patch_size: int = 2, + in_channels: int = 16, + out_channels: Optional[int] = None, + hidden_size: int = 3360, + num_layers: int = 32, + num_double_stream_layers: int = 8, + num_refiner_layers: int = 2, + num_attention_heads: int = 28, + num_kv_heads: int = 7, + multiple_of: int = 256, + ffn_dim_multiplier: Optional[float] = None, + norm_eps: float = 1e-5, + axes_dim_rope: Tuple[int, int, int] = (40, 40, 40), + axes_lens: Tuple[int, int, int] = (2048, 1664, 1664), + instruction_feat_dim: int = 4096, + timestep_scale: float = 1000.0, + image_model=None, + device=None, dtype=None, operations=None, + ): + super().__init__() + + self.patch_size = patch_size + self.out_channels = out_channels or in_channels + self.hidden_size = hidden_size + self.dtype = dtype + + self.rope_embedder = OmniGen2RotaryPosEmbed( + theta=10000, + axes_dim=axes_dim_rope, + axes_lens=axes_lens, + patch_size=patch_size, + ) + + self.x_embedder = operations.Linear(patch_size * patch_size * in_channels, hidden_size, dtype=dtype, device=device) + self.ref_image_patch_embedder = operations.Linear(patch_size * patch_size * in_channels, hidden_size, dtype=dtype, device=device) + + self.time_caption_embed = Lumina2CombinedTimestepCaptionEmbedding( + hidden_size=hidden_size, + text_feat_dim=instruction_feat_dim, + norm_eps=norm_eps, + timestep_scale=timestep_scale, dtype=dtype, device=device, operations=operations + ) + + self.noise_refiner = nn.ModuleList([ + OmniGen2TransformerBlock(hidden_size, num_attention_heads, num_kv_heads, multiple_of, ffn_dim_multiplier, norm_eps, modulation=True, dtype=dtype, device=device, operations=operations) + for _ in range(num_refiner_layers) + ]) + + self.ref_image_refiner = nn.ModuleList([ + OmniGen2TransformerBlock(hidden_size, num_attention_heads, num_kv_heads, multiple_of, ffn_dim_multiplier, norm_eps, modulation=True, dtype=dtype, device=device, operations=operations) + for _ in range(num_refiner_layers) + ]) + + self.context_refiner = nn.ModuleList([ + OmniGen2TransformerBlock(hidden_size, num_attention_heads, num_kv_heads, multiple_of, ffn_dim_multiplier, norm_eps, modulation=False, dtype=dtype, device=device, operations=operations) + for _ in range(num_refiner_layers) + ]) + + self.double_stream_layers = nn.ModuleList([ + BooguDoubleStreamBlock(hidden_size, num_attention_heads, num_kv_heads, multiple_of, ffn_dim_multiplier, norm_eps, dtype=dtype, device=device, operations=operations) + for _ in range(num_double_stream_layers) + ]) + + self.single_stream_layers = nn.ModuleList([ + OmniGen2TransformerBlock(hidden_size, num_attention_heads, num_kv_heads, multiple_of, ffn_dim_multiplier, norm_eps, modulation=True, dtype=dtype, device=device, operations=operations) + for _ in range(num_layers) + ]) + + self.norm_out = LuminaLayerNormContinuous( + embedding_dim=hidden_size, + conditioning_embedding_dim=min(hidden_size, 1024), + elementwise_affine=False, + eps=1e-6, + out_dim=patch_size * patch_size * self.out_channels, dtype=dtype, device=device, operations=operations + ) + + self.image_index_embedding = nn.Parameter(torch.empty(5, hidden_size, device=device, dtype=dtype)) + + # Patchify/refine helpers are identical to OmniGen2; reuse via bound methods. + flat_and_pad_to_seq = comfy.ldm.omnigen.omnigen2.OmniGen2Transformer2DModel.flat_and_pad_to_seq + img_patch_embed_and_refine = comfy.ldm.omnigen.omnigen2.OmniGen2Transformer2DModel.img_patch_embed_and_refine + + def forward(self, x, timesteps, context, num_tokens, ref_latents=None, attention_mask=None, transformer_options={}, **kwargs): + B, C, H, W = x.shape + hidden_states = comfy.ldm.common_dit.pad_to_patch_size(x, (self.patch_size, self.patch_size)) + _, _, H_padded, W_padded = hidden_states.shape + timestep = 1.0 - timesteps + text_hidden_states = context + text_attention_mask = attention_mask + ref_image_hidden_states = ref_latents + device = hidden_states.device + + temb, text_hidden_states = self.time_caption_embed(timestep, text_hidden_states, hidden_states[0].dtype) + + ( + hidden_states, ref_image_hidden_states, + img_mask, ref_img_mask, + l_effective_ref_img_len, l_effective_img_len, + ref_img_sizes, img_sizes, + ) = self.flat_and_pad_to_seq(hidden_states, ref_image_hidden_states) + + ( + context_rotary_emb, ref_img_rotary_emb, noise_rotary_emb, + rotary_emb, encoder_seq_lengths, seq_lengths, + ) = self.rope_embedder( + hidden_states.shape[0], text_hidden_states.shape[1], [num_tokens] * text_hidden_states.shape[0], + l_effective_ref_img_len, l_effective_img_len, + ref_img_sizes, img_sizes, device, + ) + + for layer in self.context_refiner: + text_hidden_states = layer(text_hidden_states, text_attention_mask, context_rotary_emb, transformer_options=transformer_options) + + img_len = hidden_states.shape[1] + combined_img_hidden_states = self.img_patch_embed_and_refine( + hidden_states, ref_image_hidden_states, + img_mask, ref_img_mask, + noise_rotary_emb, ref_img_rotary_emb, + l_effective_ref_img_len, l_effective_img_len, + temb, + transformer_options=transformer_options, + ) + + # Double-stream stage: the image self-attention only sees the [ref ; noise] tokens, + # which sit after the instruction tokens in the joint rope. + L_instruct = text_hidden_states.shape[1] + combined_img_rotary_emb = rotary_emb[:, L_instruct:] + for layer in self.double_stream_layers: + combined_img_hidden_states, text_hidden_states = layer( + combined_img_hidden_states, text_hidden_states, + rotary_emb, combined_img_rotary_emb, temb, + joint_attention_mask=None, img_attention_mask=None, + transformer_options=transformer_options, + ) + + hidden_states = torch.cat([text_hidden_states, combined_img_hidden_states], dim=1) + + for layer in self.single_stream_layers: + hidden_states = layer(hidden_states, None, rotary_emb, temb, transformer_options=transformer_options) + + hidden_states = self.norm_out(hidden_states, temb) + + p = self.patch_size + output = rearrange(hidden_states[:, -img_len:], 'b (h w) (p1 p2 c) -> b c (h p1) (w p2)', h=H_padded // p, w=W_padded // p, p1=p, p2=p)[:, :, :H, :W] + + return -output diff --git a/comfy/ldm/omnigen/omnigen2.py b/comfy/ldm/omnigen/omnigen2.py index e9ca5229d..b8da4cf39 100644 --- a/comfy/ldm/omnigen/omnigen2.py +++ b/comfy/ldm/omnigen/omnigen2.py @@ -22,7 +22,7 @@ def apply_rotary_emb(x, freqs_cis): def swiglu(x: torch.Tensor, y: torch.Tensor) -> torch.Tensor: - return F.silu(x) * y + return F.silu(x, inplace=True).mul_(y) class TimestepEmbedding(nn.Module): diff --git a/comfy/model_base.py b/comfy/model_base.py index d143dc06f..f49da50ae 100644 --- a/comfy/model_base.py +++ b/comfy/model_base.py @@ -54,6 +54,7 @@ import comfy.ldm.pixeldit.model import comfy.ldm.pixeldit.pid import comfy.ldm.ace.model import comfy.ldm.omnigen.omnigen2 +import comfy.ldm.boogu.model import comfy.ldm.qwen_image.model import comfy.ldm.ideogram4.model import comfy.ldm.kandinsky5.model @@ -2103,6 +2104,11 @@ class Omnigen2(BaseModel): out['ref_latents'] = list([1, 16, sum(map(lambda a: math.prod(a.size()), ref_latents)) // 16]) return out +class Boogu(Omnigen2): + def __init__(self, model_config, model_type=ModelType.FLOW, device=None): + super(Omnigen2, self).__init__(model_config, model_type, device=device, unet_model=comfy.ldm.boogu.model.BooguTransformer2DModel) + self.memory_usage_factor_conds = ("ref_latents",) + class QwenImage(BaseModel): def __init__(self, model_config, model_type=ModelType.FLUX, device=None): super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.qwen_image.model.QwenImageTransformer2DModel) diff --git a/comfy/model_detection.py b/comfy/model_detection.py index 7d0cab308..b773f0393 100644 --- a/comfy/model_detection.py +++ b/comfy/model_detection.py @@ -761,6 +761,16 @@ def detect_unet_config(state_dict, key_prefix, metadata=None): return dit_config + if '{}double_stream_layers.0.img_instruct_attn.processor.img_to_q.weight'.format(key_prefix) in state_dict_keys: # Boogu-Image (OmniGen2 derivative + dual-stream stage) + dit_config = {} + dit_config["image_model"] = "boogu" + dit_config["hidden_size"] = state_dict['{}x_embedder.weight'.format(key_prefix)].shape[0] + dit_config["num_layers"] = count_blocks(state_dict_keys, '{}single_stream_layers.'.format(key_prefix) + '{}.') + dit_config["num_double_stream_layers"] = count_blocks(state_dict_keys, '{}double_stream_layers.'.format(key_prefix) + '{}.') + dit_config["num_refiner_layers"] = count_blocks(state_dict_keys, '{}noise_refiner.'.format(key_prefix) + '{}.') + dit_config["instruction_feat_dim"] = state_dict['{}time_caption_embed.caption_embedder.0.weight'.format(key_prefix)].shape[0] + return dit_config + if '{}time_caption_embed.timestep_embedder.linear_1.bias'.format(key_prefix) in state_dict_keys: # Omnigen2 dit_config = {} dit_config["image_model"] = "omnigen2" diff --git a/comfy/sd.py b/comfy/sd.py index 348fe4958..d9b1c0553 100644 --- a/comfy/sd.py +++ b/comfy/sd.py @@ -68,6 +68,7 @@ import comfy.text_encoders.ace15 import comfy.text_encoders.longcat_image import comfy.text_encoders.qwen35 import comfy.text_encoders.qwen3vl +import comfy.text_encoders.boogu import comfy.text_encoders.ernie import comfy.text_encoders.gemma4 import comfy.text_encoders.cogvideo @@ -1301,6 +1302,7 @@ class CLIPType(Enum): LENS = 28 PIXELDIT = 29 IDEOGRAM4 = 30 + BOOGU = 31 @@ -1622,6 +1624,10 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip clip_data[0] = comfy.utils.state_dict_prefix_replace(clip_data[0], {"model.language_model.": "model.", "model.visual.": "visual.", "lm_head.": "model.lm_head."}) clip_target.clip = comfy.text_encoders.ideogram4.te_qwen3vl(**llama_detect(clip_data)) clip_target.tokenizer = comfy.text_encoders.ideogram4.Ideogram4Qwen3VLTokenizer + elif clip_type == CLIPType.BOOGU and te_model == TEModel.QWEN3VL_8B: # Boogu-Image: full Qwen3-VL-8B, last hidden state, no-think template. + clip_data[0] = comfy.utils.state_dict_prefix_replace(clip_data[0], {"model.language_model.": "model.", "model.visual.": "visual.", "lm_head.": "model.lm_head."}) + clip_target.clip = comfy.text_encoders.boogu.te(**llama_detect(clip_data)) + clip_target.tokenizer = comfy.text_encoders.boogu.BooguTokenizer elif clip_type in (CLIPType.FLUX, CLIPType.FLUX2): # Flux2 Klein reuses the Qwen3-VL LM (3-layer tap -> 12288); visual unused. klein_model_type = "qwen3_8b" if te_model == TEModel.QWEN3VL_8B else "qwen3_4b" clip_target.clip = comfy.text_encoders.flux.klein_te(**llama_detect(clip_data), model_type=klein_model_type) diff --git a/comfy/supported_models.py b/comfy/supported_models.py index 3be935577..d78b6ae87 100644 --- a/comfy/supported_models.py +++ b/comfy/supported_models.py @@ -25,6 +25,7 @@ import comfy.text_encoders.hunyuan_image import comfy.text_encoders.kandinsky5 import comfy.text_encoders.z_image import comfy.text_encoders.ideogram4 +import comfy.text_encoders.boogu import comfy.text_encoders.anima import comfy.text_encoders.ace15 import comfy.text_encoders.longcat_image @@ -1758,6 +1759,27 @@ class Omnigen2(supported_models_base.BASE): hunyuan_detect = comfy.text_encoders.hunyuan_video.llama_detect(state_dict, "{}qwen25_3b.transformer.".format(pref)) return supported_models_base.ClipTarget(comfy.text_encoders.omnigen2.Omnigen2Tokenizer, comfy.text_encoders.omnigen2.te(**hunyuan_detect)) +class Boogu(Omnigen2): + unet_config = { + "image_model": "boogu", + } + + sampling_settings = { + "multiplier": 1.0, + "shift": 3.16, + } + + memory_usage_factor = 1.95 #TODO + + def get_model(self, state_dict, prefix="", device=None): + out = model_base.Boogu(self, device=device) + return out + + def clip_target(self, state_dict={}): + pref = self.text_encoder_key_prefix[0] + hunyuan_detect = comfy.text_encoders.hunyuan_video.llama_detect(state_dict, "{}qwen3vl_8b.transformer.".format(pref)) + return supported_models_base.ClipTarget(comfy.text_encoders.boogu.BooguTokenizer, comfy.text_encoders.boogu.te(**hunyuan_detect)) + class Ideogram4(supported_models_base.BASE): unet_config = { "image_model": "ideogram4", @@ -2300,6 +2322,7 @@ models = [ ACEStep, ACEStep15, Omnigen2, + Boogu, QwenImage, Ideogram4, Flux2, diff --git a/comfy/text_encoders/boogu.py b/comfy/text_encoders/boogu.py new file mode 100644 index 000000000..d9de92f10 --- /dev/null +++ b/comfy/text_encoders/boogu.py @@ -0,0 +1,58 @@ +"""Boogu-Image text encoder: full Qwen3-VL-8B, last hidden state (4096-dim). + +Boogu uses the final hidden state of Qwen3-VL as the per-token instruction feature +(num_instruction_feature_layers=1, reduce_type=mean -> just the last layer). +The model itself is the standard Qwen3-VL TE, only the chat template differs +(a fixed system prompt and no block). +""" + +import comfy.text_encoders.qwen3vl +from comfy import sd1_clip + + +# System prompts from the reference pipeline (pipeline_boogu.py). +# T2I (non-empty instruction, no image) uses the helpful-assistant prompt +# everything else (the CFG negative / "drop" condition, and any image case) uses the TI2I "describe" prompt. +BOOGU_T2I_SYSTEM = "You are a helpful assistant that generates high-quality images based on user instructions. The instructions are as follows." +BOOGU_DROP_SYSTEM = "Describe the key features of the input image (color, shape, size, texture, objects, background), then explain how the user's text instruction should alter or modify the image. Generate a new image that meets the user's requirements while maintaining consistency with the original input where appropriate." + + +class BooguTokenizer(comfy.text_encoders.qwen3vl.Qwen3VLTokenizer): + def __init__(self, embedding_directory=None, tokenizer_data={}): + super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, model_type="qwen3vl_8b") + # apply_chat_template without add_generation_prompt + self.llama_template = "<|im_start|>system\n" + BOOGU_T2I_SYSTEM + "<|im_end|>\n<|im_start|>user\n{}<|im_end|>\n" + self.llama_template_images = "<|im_start|>system\n" + BOOGU_DROP_SYSTEM + "<|im_end|>\n<|im_start|>user\n<|vision_start|><|image_pad|><|vision_end|>{}<|im_end|>\n" + # Reference SYSTEM_PROMPT_DROP: used for the empty negative/uncond instruction. + self.llama_template_drop = "<|im_start|>system\n" + BOOGU_DROP_SYSTEM + "<|im_end|>\n<|im_start|>user\n{}<|im_end|>\n" + + def tokenize_with_weights(self, text, return_word_ids=False, llama_template=None, images=[], prevent_empty_text=False, thinking=True, **kwargs): + if llama_template is None and len(images) == 0 and text.strip() == "": + llama_template = self.llama_template_drop + # Boogu conditions on the no-think template; thinking=True drops the empty block qwen3vl adds by default. + return super().tokenize_with_weights(text, return_word_ids=return_word_ids, llama_template=llama_template, images=images, prevent_empty_text=prevent_empty_text, thinking=thinking, **kwargs) + + +class BooguQwen3VLClipModel(comfy.text_encoders.qwen3vl.Qwen3VLClipModel): + def __init__(self, device="cpu", dtype=None, attention_mask=True, model_options={}, model_type="qwen3vl_8b"): + super().__init__(device=device, dtype=dtype, attention_mask=attention_mask, model_options=model_options, model_type=model_type) + # apply the final RMSNorm to the tapped last layer + self.layer_norm_hidden_state = True + + +class BooguTEModel(sd1_clip.SD1ClipModel): + def __init__(self, device="cpu", dtype=None, model_options={}): + clip_model = lambda **kw: BooguQwen3VLClipModel(**kw, model_type="qwen3vl_8b") + super().__init__(device=device, dtype=dtype, name="qwen3vl_8b", clip_model=clip_model, model_options=model_options) + + +def te(dtype_llama=None, llama_quantization_metadata=None): + class BooguTEModel_(BooguTEModel): + def __init__(self, device="cpu", dtype=None, model_options={}): + if dtype_llama is not None: + dtype = dtype_llama + if llama_quantization_metadata is not None: + model_options = model_options.copy() + model_options["quantization_metadata"] = llama_quantization_metadata + super().__init__(device=device, dtype=dtype, model_options=model_options) + return BooguTEModel_ diff --git a/comfy_extras/nodes_boogu.py b/comfy_extras/nodes_boogu.py new file mode 100644 index 000000000..590623e12 --- /dev/null +++ b/comfy_extras/nodes_boogu.py @@ -0,0 +1,96 @@ +import math + +import node_helpers +import comfy.utils +from typing_extensions import override +from comfy_api.latest import ComfyExtension, io + + +class TextEncodeBooguEdit(io.ComfyNode): + """Boogu-Image Edit conditioning. + + The edit image is used twice, matching the reference pipeline: + - Qwen3-VL vision tokens (instruction understanding) -> positive only + - VAE reference latent (image identity) -> positive and negative + The ref latent is in both conds so it cancels under CFG (identity preserved); + the vision tokens are only in the positive so CFG amplifies the instruction. + The tokenizer selects the right system prompt automatically (image -> TI2I, + empty negative -> DROP), so no template plumbing is needed here. + """ + + @classmethod + def define_schema(cls): + return io.Schema( + node_id="TextEncodeBooguEdit", + category="model/conditioning/boogu", + inputs=[ + io.Clip.Input("clip"), + io.String.Input("prompt", multiline=True, dynamic_prompts=True), + io.Vae.Input("vae"), + io.Autogrow.Input( + "images", + template=io.Autogrow.TemplateNames( + io.Image.Input("image"), + names=[f"image_{i}" for i in range(1, 17)], + min=1, + ), + tooltip="Reference image(s) to edit. Boogu focuses on one reference per sample; more are allowed.", + ), + ], + outputs=[ + io.Conditioning.Output(display_name="positive"), + io.Conditioning.Output(display_name="negative"), + ], + ) + + @classmethod + def execute(cls, clip, prompt, vae=None, images: io.Autogrow.Type = None) -> io.NodeOutput: + ref_latents = [] + images_vl = [] + + images = images or {} + for name in sorted(images, key=lambda n: int(n.rsplit("_", 1)[-1])): + image = images[name] + if image is None: + continue + samples = image.movedim(-1, 1) + + # Vision tower input: the reference caps the VLM image at 384x384 + # (max_vlm_input_pil_pixels in pipeline_boogu.py). + total = int(384 * 384) + scale_by = math.sqrt(total / (samples.shape[3] * samples.shape[2])) + width = round(samples.shape[3] * scale_by) + height = round(samples.shape[2] * scale_by) + s = comfy.utils.common_upscale(samples, width, height, "area", "disabled") + images_vl.append(s.movedim(1, -1)[:, :, :, :3]) + + # Reference latent: align to 16 px (VAE /8 * patch_size 2). + if vae is not None: + total = int(1024 * 1024) + scale_by = math.sqrt(total / (samples.shape[3] * samples.shape[2])) + width = round(samples.shape[3] * scale_by / 16.0) * 16 + height = round(samples.shape[2] * scale_by / 16.0) * 16 + s = comfy.utils.common_upscale(samples, width, height, "area", "disabled") + ref_latents.append(vae.encode(s.movedim(1, -1)[:, :, :, :3])) + + # positive: instruction + vision tokens; negative: empty (no vision). Ref latent on both. + positive = clip.encode_from_tokens_scheduled(clip.tokenize(prompt, images=images_vl)) + negative = clip.encode_from_tokens_scheduled(clip.tokenize("")) + + if len(ref_latents) > 0: + positive = node_helpers.conditioning_set_values(positive, {"reference_latents": ref_latents}, append=True) + negative = node_helpers.conditioning_set_values(negative, {"reference_latents": ref_latents}, append=True) + + return io.NodeOutput(positive, negative) + + +class BooguExtension(ComfyExtension): + @override + async def get_node_list(self) -> list[type[io.ComfyNode]]: + return [ + TextEncodeBooguEdit, + ] + + +async def comfy_entrypoint() -> BooguExtension: + return BooguExtension() diff --git a/nodes.py b/nodes.py index bb4649478..0b3fdab63 100644 --- a/nodes.py +++ b/nodes.py @@ -969,7 +969,7 @@ class CLIPLoader: @classmethod def INPUT_TYPES(s): return {"required": { "clip_name": (folder_paths.get_filename_list("text_encoders"), ), - "type": (["stable_diffusion", "stable_cascade", "sd3", "stable_audio", "mochi", "ltxv", "pixart", "cosmos", "lumina2", "wan", "hidream", "chroma", "ace", "omnigen2", "qwen_image", "hunyuan_image", "flux2", "ovis", "longcat_image", "cogvideox", "lens", "pixeldit", "ideogram4"], ), + "type": (["stable_diffusion", "stable_cascade", "sd3", "stable_audio", "mochi", "ltxv", "pixart", "cosmos", "lumina2", "wan", "hidream", "chroma", "ace", "omnigen2", "qwen_image", "hunyuan_image", "flux2", "ovis", "longcat_image", "cogvideox", "lens", "pixeldit", "ideogram4", "boogu"], ), }, "optional": { "device": (["default", "cpu"], {"advanced": True}), @@ -2425,6 +2425,7 @@ async def init_builtin_extra_nodes(): "nodes_tcfg.py", "nodes_context_windows.py", "nodes_qwen.py", + "nodes_boogu.py", "nodes_chroma_radiance.py", "nodes_pid.py", "nodes_model_patch.py", From 52257bb435162ff345fcc8cfde99c52c8923589d Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Wed, 17 Jun 2026 15:42:29 -0700 Subject: [PATCH 05/65] Add negative prompt to boogu edit node and set min images to 0. (#14529) --- comfy/supported_models.py | 2 +- comfy_extras/nodes_boogu.py | 7 ++++--- 2 files changed, 5 insertions(+), 4 deletions(-) diff --git a/comfy/supported_models.py b/comfy/supported_models.py index d78b6ae87..cc05908ee 100644 --- a/comfy/supported_models.py +++ b/comfy/supported_models.py @@ -1769,7 +1769,7 @@ class Boogu(Omnigen2): "shift": 3.16, } - memory_usage_factor = 1.95 #TODO + memory_usage_factor = 2.15 def get_model(self, state_dict, prefix="", device=None): out = model_base.Boogu(self, device=device) diff --git a/comfy_extras/nodes_boogu.py b/comfy_extras/nodes_boogu.py index 590623e12..f3951c290 100644 --- a/comfy_extras/nodes_boogu.py +++ b/comfy_extras/nodes_boogu.py @@ -26,13 +26,14 @@ class TextEncodeBooguEdit(io.ComfyNode): inputs=[ io.Clip.Input("clip"), io.String.Input("prompt", multiline=True, dynamic_prompts=True), + io.String.Input("negative_prompt", multiline=True, dynamic_prompts=True, advanced=True), io.Vae.Input("vae"), io.Autogrow.Input( "images", template=io.Autogrow.TemplateNames( io.Image.Input("image"), names=[f"image_{i}" for i in range(1, 17)], - min=1, + min=0, ), tooltip="Reference image(s) to edit. Boogu focuses on one reference per sample; more are allowed.", ), @@ -44,7 +45,7 @@ class TextEncodeBooguEdit(io.ComfyNode): ) @classmethod - def execute(cls, clip, prompt, vae=None, images: io.Autogrow.Type = None) -> io.NodeOutput: + def execute(cls, clip, prompt, negative_prompt, vae=None, images: io.Autogrow.Type = None) -> io.NodeOutput: ref_latents = [] images_vl = [] @@ -75,7 +76,7 @@ class TextEncodeBooguEdit(io.ComfyNode): # positive: instruction + vision tokens; negative: empty (no vision). Ref latent on both. positive = clip.encode_from_tokens_scheduled(clip.tokenize(prompt, images=images_vl)) - negative = clip.encode_from_tokens_scheduled(clip.tokenize("")) + negative = clip.encode_from_tokens_scheduled(clip.tokenize(negative_prompt)) if len(ref_latents) > 0: positive = node_helpers.conditioning_set_values(positive, {"reference_latents": ref_latents}, append=True) From 191a75a2cdc021236fb7cbd4631511246b79a8c6 Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Thu, 18 Jun 2026 02:54:53 +0300 Subject: [PATCH 06/65] [Partner Nodes] feat(Kling): add support for Kling V3-Turbo model (#14528) --- comfy_api_nodes/apis/kling.py | 56 ++++++++++++++++ comfy_api_nodes/nodes_kling.py | 116 ++++++++++++++++++++++++++++++--- 2 files changed, 163 insertions(+), 9 deletions(-) diff --git a/comfy_api_nodes/apis/kling.py b/comfy_api_nodes/apis/kling.py index fe0f97cb3..2c98c23b7 100644 --- a/comfy_api_nodes/apis/kling.py +++ b/comfy_api_nodes/apis/kling.py @@ -149,3 +149,59 @@ class MotionControlRequest(BaseModel): character_orientation: str = Field(...) mode: str = Field(..., description="'pro' or 'std'") model_name: str = Field(...) + + +class Kling3TurboSettings(BaseModel): + resolution: str = Field("720p", description="'720p' or '1080p'") + aspect_ratio: str | None = Field(None, description="'16:9'/'9:16'/'1:1'; text-to-video only") + duration: int = Field(5, description="3-15 second") + + +class Kling3TurboText2VideoRequest(BaseModel): + prompt: str = Field(..., description="<=3072 chars; may use multi-shot 'shot n, m, words; ...'") + settings: Kling3TurboSettings | None = Field(None) + + +class Kling3TurboContent(BaseModel): + type: str = Field(..., description="'prompt' or 'first_frame'") + text: str | None = Field(None, description="for type=prompt; <=2500 chars") + url: str | None = Field(None, description="for type=first_frame") + + +class Kling3TurboImage2VideoRequest(BaseModel): + contents: list[Kling3TurboContent] = Field(..., description="prompt + first_frame materials") + settings: Kling3TurboSettings | None = Field(None) + + +class Kling3TurboCreateData(BaseModel): + id: str | None = Field(None, description="Task ID") + status: str | None = Field(None) + message: str | None = Field(None) + + +class Kling3TurboCreateResponse(BaseModel): + code: int | None = Field(None) + message: str | None = Field(None) + request_id: str | None = Field(None) + data: Kling3TurboCreateData | None = Field(None) + + +class Kling3TurboOutput(BaseModel): + type: str | None = Field(None, description="'video', 'image', 'audio', ...") + id: str | None = Field(None) + url: str | None = Field(None) + duration: str | None = Field(None) + + +class Kling3TurboTaskData(BaseModel): + id: str | None = Field(None) + status: str | None = Field(None, description="submitted | processing | succeeded | failed") + message: str | None = Field(None) + outputs: list[Kling3TurboOutput] | None = Field(None) + + +class Kling3TurboQueryResponse(BaseModel): + code: int | None = Field(None) + message: str | None = Field(None) + request_id: str | None = Field(None) + data: list[Kling3TurboTaskData] | None = Field(None) diff --git a/comfy_api_nodes/nodes_kling.py b/comfy_api_nodes/nodes_kling.py index c81d3503d..b27de2549 100644 --- a/comfy_api_nodes/nodes_kling.py +++ b/comfy_api_nodes/nodes_kling.py @@ -60,6 +60,12 @@ from comfy_api_nodes.apis.kling import ( OmniProImageRequest, OmniProReferences2VideoRequest, OmniProText2VideoRequest, + Kling3TurboSettings, + Kling3TurboText2VideoRequest, + Kling3TurboContent, + Kling3TurboImage2VideoRequest, + Kling3TurboCreateResponse, + Kling3TurboQueryResponse, TaskStatusResponse, TextToVideoWithAudioRequest, ) @@ -2847,6 +2853,67 @@ class MotionControl(IO.ComfyNode): return IO.NodeOutput(await download_url_to_video_output(final_response.data.task_result.videos[0].url)) +def build_turbo_shot_prompt(multi_prompt: list[MultiPromptEntry]) -> str: + """Render storyboard entries into the Turbo multi-shot prompt 'shot n, m, words; ...'.""" + return "; ".join(f"shot {i}, {int(e.duration)}, {e.prompt}" for i, e in enumerate(multi_prompt, 1)) + ";" + + +def _turbo_video_url(response: Kling3TurboQueryResponse) -> str: + """Extract the result video URL from a /tasks response (data[].outputs[] where type == 'video').""" + task = response.data[0] if response.data else None + if task and task.outputs: + for output in task.outputs: + if output.type == "video" and output.url: + return output.url + raise RuntimeError(f"Kling 3.0 Turbo task finished without a video output: {response.model_dump()}") + + +async def execute_kling_turbo( + cls: type[IO.ComfyNode], + *, + prompt: str, + resolution: str, + aspect_ratio: str, + duration: int, + start_frame: torch.Tensor | None, +) -> IO.NodeOutput: + """Create + poll a Kling 3.0 Turbo task. Image-to-video when start_frame is given, else text-to-video.""" + if start_frame is not None: + validate_image_dimensions(start_frame, min_width=300, min_height=300) + validate_image_aspect_ratio(start_frame, (1, 2.5), (2.5, 1)) + contents = [Kling3TurboContent(type="first_frame", url=tensor_to_base64_string(start_frame))] + if prompt: + contents.insert(0, Kling3TurboContent(type="prompt", text=prompt)) + create = await sync_op( + cls, + ApiEndpoint(path="/proxy/kling/image-to-video/kling-3.0-turbo", method="POST"), + response_model=Kling3TurboCreateResponse, + data=Kling3TurboImage2VideoRequest( + contents=contents, + settings=Kling3TurboSettings(resolution=resolution, duration=duration), # i2v: no aspect_ratio + ), + ) + else: + create = await sync_op( + cls, + ApiEndpoint(path="/proxy/kling/text-to-video/kling-3.0-turbo", method="POST"), + response_model=Kling3TurboCreateResponse, + data=Kling3TurboText2VideoRequest( + prompt=prompt, + settings=Kling3TurboSettings(resolution=resolution, aspect_ratio=aspect_ratio, duration=duration), + ), + ) + if not (create.data and create.data.id): + raise RuntimeError(f"Kling 3.0 Turbo create failed. Code: {create.code}, Message: {create.message}") + final_response = await poll_op( + cls, + ApiEndpoint(path="/proxy/kling/tasks", query_params={"task_ids": create.data.id}), + response_model=Kling3TurboQueryResponse, + status_extractor=lambda r: (r.data[0].status if r.data else None), + ) + return IO.NodeOutput(await download_url_to_video_output(_turbo_video_url(final_response))) + + class KlingVideoNode(IO.ComfyNode): @classmethod @@ -2884,7 +2951,11 @@ class KlingVideoNode(IO.ComfyNode): ], tooltip="Generate a series of video segments with individual prompts and durations.", ), - IO.Boolean.Input("generate_audio", default=True), + IO.Boolean.Input( + "generate_audio", + default=True, + tooltip="'kling-3.0-turbo' always generates native audio, so the audio toggle is ignored.", + ), IO.DynamicCombo.Input( "model", options=[ @@ -2899,6 +2970,17 @@ class KlingVideoNode(IO.ComfyNode): ), ], ), + IO.DynamicCombo.Option( + "kling-3.0-turbo", + [ + IO.Combo.Input("resolution", options=["1080p", "720p"], default="720p"), + IO.Combo.Input( + "aspect_ratio", + options=["16:9", "9:16", "1:1"], + tooltip="Ignored in image-to-video mode.", + ), + ], + ), ], tooltip="Model and generation settings.", ), @@ -2930,6 +3012,7 @@ class KlingVideoNode(IO.ComfyNode): price_badge=IO.PriceBadge( depends_on=IO.PriceBadgeDepends( widgets=[ + "model", "model.resolution", "generate_audio", "multi_shot", @@ -2944,14 +3027,7 @@ class KlingVideoNode(IO.ComfyNode): ), expr=""" ( - $rates := { - "4k": {"off": 0.42, "on": 0.42}, - "1080p": {"off": 0.112, "on": 0.168}, - "720p": {"off": 0.084, "on": 0.126} - }; $res := $lookup(widgets, "model.resolution"); - $audio := widgets.generate_audio ? "on" : "off"; - $rate := $lookup($lookup($rates, $res), $audio); $ms := widgets.multi_shot; $isSb := $ms != "disabled"; $n := $isSb ? $number($substring($ms, 0, 1)) : 0; @@ -2962,7 +3038,18 @@ class KlingVideoNode(IO.ComfyNode): $d5 := $n >= 5 ? $lookup(widgets, "multi_shot.storyboard_5_duration") : 0; $d6 := $n >= 6 ? $lookup(widgets, "multi_shot.storyboard_6_duration") : 0; $dur := $isSb ? $d1 + $d2 + $d3 + $d4 + $d5 + $d6 : $lookup(widgets, "multi_shot.duration"); - {"type":"usd","usd": $rate * $dur} + widgets.model = "kling-3.0-turbo" + ? {"type":"usd","usd": ($res = "1080p" ? 0.14 : 0.112) * $dur} + : ( + $rates := { + "4k": {"off": 0.42, "on": 0.42}, + "1080p": {"off": 0.112, "on": 0.168}, + "720p": {"off": 0.084, "on": 0.126} + }; + $audio := widgets.generate_audio ? "on" : "off"; + $rate := $lookup($lookup($rates, $res), $audio); + {"type":"usd","usd": $rate * $dur} + ) ) """, ), @@ -3015,6 +3102,17 @@ class KlingVideoNode(IO.ComfyNode): duration = multi_shot["duration"] validate_string(multi_shot["prompt"], min_length=1, max_length=2500) + if model["model"] == "kling-3.0-turbo": + turbo_prompt = build_turbo_shot_prompt(multi_prompt_list) if custom_multi_shot else multi_shot["prompt"] + return await execute_kling_turbo( + cls, + prompt=turbo_prompt, + resolution=model["resolution"], + aspect_ratio=model["aspect_ratio"], + duration=duration, + start_frame=start_frame, + ) + if start_frame is not None: validate_image_dimensions(start_frame, min_width=300, min_height=300) validate_image_aspect_ratio(start_frame, (1, 2.5), (2.5, 1)) From f2270f070a372898ca748299a51dcba5bba0ba53 Mon Sep 17 00:00:00 2001 From: Jedrzej Kosinski Date: Wed, 17 Jun 2026 19:35:05 -0700 Subject: [PATCH 07/65] feat: add enable_telemetry CLI feature flag (#14530) --- comfy_api/feature_flags.py | 5 +++++ 1 file changed, 5 insertions(+) diff --git a/comfy_api/feature_flags.py b/comfy_api/feature_flags.py index adb5a3144..0f30608a9 100644 --- a/comfy_api/feature_flags.py +++ b/comfy_api/feature_flags.py @@ -25,6 +25,11 @@ CLI_FEATURE_FLAG_REGISTRY: dict[str, FeatureFlagInfo] = { "default": False, "description": "Show the sign-in button in the frontend even when not signed in", }, + "enable_telemetry": { + "type": "bool", + "default": False, + "description": "Signal the frontend that telemetry collection is enabled", + }, } From 8483c215dc4b3958a75a4b094755446e4de63e23 Mon Sep 17 00:00:00 2001 From: "Daxiong (Lin)" Date: Thu, 18 Jun 2026 17:24:05 +0800 Subject: [PATCH 08/65] Update ComfyUI Desktop to Comfy Desktop for consistent product naming (#14533) --- README.md | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/README.md b/README.md index c75353d36..bcec86377 100644 --- a/README.md +++ b/README.md @@ -140,7 +140,7 @@ ComfyUI follows a weekly release cycle targeting Monday but this regularly chang - Commits outside of the stable release tags may be very unstable and break many custom nodes. - Serves as the foundation for the desktop release -2. **[ComfyUI Desktop](https://github.com/Comfy-Org/Comfy-Desktop)** +2. **[Comfy Desktop](https://github.com/Comfy-Org/Comfy-Desktop)** - Builds a new release using the latest stable core version 3. **[ComfyUI Frontend](https://github.com/Comfy-Org/ComfyUI_frontend)** From 16514da2e75fe73e592675ef1740eb3c07e4f80c Mon Sep 17 00:00:00 2001 From: Comfy Org PR Bot Date: Thu, 18 Jun 2026 18:27:53 +0900 Subject: [PATCH 09/65] chore(openapi): sync shared API contract from cloud@d10ff72 (#14518) --- openapi.yaml | 195 ++++++++++++++++++++++++++++++++++++++++++++++++--- 1 file changed, 187 insertions(+), 8 deletions(-) diff --git a/openapi.yaml b/openapi.yaml index 82ff5b003..2446e64e4 100644 --- a/openapi.yaml +++ b/openapi.yaml @@ -673,6 +673,35 @@ components: - created_at - updated_at type: object + JobsCancelRequest: + additionalProperties: false + description: Request to cancel multiple jobs by ID. + properties: + job_ids: + description: Job identifiers (UUIDs) to cancel. + items: + format: uuid + type: string + maxItems: 100 + minItems: 1 + type: array + required: + - job_ids + type: object + JobsCancelResponse: + description: Response for POST /api/jobs/cancel. + properties: + cancelled: + description: | + Job IDs for which a cancel event was successfully dispatched by this + call. Jobs already in a terminal or cancelling state are idempotently + skipped and will not appear here. + items: + type: string + type: array + required: + - cancelled + type: object JobsListResponse: description: Paginated list of jobs for the authenticated user. properties: @@ -1006,7 +1035,7 @@ components: description: If true, clear all pending jobs from the queue type: boolean delete: - description: Array of PENDING job IDs to cancel + description: Array of job IDs to cancel; pending and running jobs transition to cancelled items: type: string type: array @@ -1822,6 +1851,83 @@ paths: summary: Update asset metadata tags: - file + /api/assets/{id}/content: + get: + description: | + Returns the binary content of an asset by ID. + + The contract is the same across runtimes — "GET this path and you + receive the asset's bytes" — but the mechanism differs: + - **Local ComfyUI** streams the bytes directly (`200`, + `application/octet-stream`). + - **Cloud** does not proxy large files; it responds `302` with a + `Location` redirect to a short-lived signed storage URL. Clients that + follow redirects (browsers, `fetch`/XHR, ``/`