From fb4739f2f5285bee9cef4232c53e5602b8390a2b Mon Sep 17 00:00:00 2001 From: kijai <40791699+kijai@users.noreply.github.com> Date: Tue, 18 Nov 2025 21:42:46 +0200 Subject: [PATCH] Support HunyuanVideo1.5 SR model --- comfy/model_base.py | 9 +++++++ comfy/supported_models.py | 27 ++++++++++++++++++++- comfy_extras/nodes_hunyuan.py | 45 +++++++++++++++++++++++++++++++++++ 3 files changed, 80 insertions(+), 1 deletion(-) diff --git a/comfy/model_base.py b/comfy/model_base.py index 16bc1b0c2..5d2b46167 100644 --- a/comfy/model_base.py +++ b/comfy/model_base.py @@ -1596,3 +1596,12 @@ class HunyuanVideo15(HunyuanVideo): out['clip_fea'] = comfy.conds.CONDRegular(clip_vision_output.last_hidden_state) return out + +class HunyuanVideo15_SR_Distilled(HunyuanImage21Refiner): + def __init__(self, model_config, model_type=ModelType.FLOW, device=None): + super().__init__(model_config, model_type, device=device) + + def extra_conds(self, **kwargs): + out = super().extra_conds(**kwargs) + out['disable_time_r'] = comfy.conds.CONDConstant(False) + return out diff --git a/comfy/supported_models.py b/comfy/supported_models.py index 3fd21bfe5..2e64b85e8 100644 --- a/comfy/supported_models.py +++ b/comfy/supported_models.py @@ -1397,6 +1397,31 @@ class HunyuanVideo15(HunyuanVideo): hunyuan_detect = comfy.text_encoders.hunyuan_video.llama_detect(state_dict, "{}qwen25_7b.transformer.".format(pref)) return supported_models_base.ClipTarget(comfy.text_encoders.hunyuan_video.HunyuanVideo15Tokenizer, comfy.text_encoders.hunyuan_image.te(**hunyuan_detect)) -models = [LotusD, Stable_Zero123, SD15_instructpix2pix, SD15, SD20, SD21UnclipL, SD21UnclipH, SDXL_instructpix2pix, SDXLRefiner, SDXL, SSD1B, KOALA_700M, KOALA_1B, Segmind_Vega, SD_X4Upscaler, Stable_Cascade_C, Stable_Cascade_B, SV3D_u, SV3D_p, SD3, StableAudio, AuraFlow, PixArtAlpha, PixArtSigma, HunyuanDiT, HunyuanDiT1, FluxInpaint, Flux, FluxSchnell, GenmoMochi, LTXV, HunyuanVideo15, HunyuanImage21Refiner, HunyuanImage21, HunyuanVideoSkyreelsI2V, HunyuanVideoI2V, HunyuanVideo, CosmosT2V, CosmosI2V, CosmosT2IPredict2, CosmosI2VPredict2, Lumina2, WAN22_T2V, WAN21_T2V, WAN21_I2V, WAN21_FunControl2V, WAN21_Vace, WAN21_Camera, WAN22_Camera, WAN22_S2V, WAN21_HuMo, WAN22_Animate, Hunyuan3Dv2mini, Hunyuan3Dv2, Hunyuan3Dv2_1, HiDream, Chroma, ChromaRadiance, ACEStep, Omnigen2, QwenImage] + +class HunyuanVideo15_SR_Distilled(HunyuanVideo): + unet_config = { + "image_model": "hunyuan_video", + "vision_in_dim": 1152, + "in_channels": 98, + } + + sampling_settings = { + "shift": 2.0, + } + memory_usage_factor = 4.0 #TODO + supported_inference_dtypes = [torch.float16, torch.bfloat16, torch.float32] + + latent_format = latent_formats.HunyuanVideo15 + + def get_model(self, state_dict, prefix="", device=None): + out = model_base.HunyuanVideo15_SR_Distilled(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, "{}qwen25_7b.transformer.".format(pref)) + return supported_models_base.ClipTarget(comfy.text_encoders.hunyuan_video.HunyuanVideo15Tokenizer, comfy.text_encoders.hunyuan_image.te(**hunyuan_detect)) + +models = [LotusD, Stable_Zero123, SD15_instructpix2pix, SD15, SD20, SD21UnclipL, SD21UnclipH, SDXL_instructpix2pix, SDXLRefiner, SDXL, SSD1B, KOALA_700M, KOALA_1B, Segmind_Vega, SD_X4Upscaler, Stable_Cascade_C, Stable_Cascade_B, SV3D_u, SV3D_p, SD3, StableAudio, AuraFlow, PixArtAlpha, PixArtSigma, HunyuanDiT, HunyuanDiT1, FluxInpaint, Flux, FluxSchnell, GenmoMochi, LTXV, HunyuanVideo15_SR_Distilled, HunyuanVideo15, HunyuanImage21Refiner, HunyuanImage21, HunyuanVideoSkyreelsI2V, HunyuanVideoI2V, HunyuanVideo, CosmosT2V, CosmosI2V, CosmosT2IPredict2, CosmosI2VPredict2, Lumina2, WAN22_T2V, WAN21_T2V, WAN21_I2V, WAN21_FunControl2V, WAN21_Vace, WAN21_Camera, WAN22_Camera, WAN22_S2V, WAN21_HuMo, WAN22_Animate, Hunyuan3Dv2mini, Hunyuan3Dv2, Hunyuan3Dv2_1, HiDream, Chroma, ChromaRadiance, ACEStep, Omnigen2, QwenImage] models += [SVD_img2vid] diff --git a/comfy_extras/nodes_hunyuan.py b/comfy_extras/nodes_hunyuan.py index dfef2e0ab..eef606eca 100644 --- a/comfy_extras/nodes_hunyuan.py +++ b/comfy_extras/nodes_hunyuan.py @@ -125,6 +125,50 @@ class HunyuanVideo15ImageToVideo(io.ComfyNode): encode = execute # TODO: remove +class HunyuanVideo15RefinerLatent(io.ComfyNode): + @classmethod + def define_schema(cls): + return io.Schema( + node_id="HunyuanVideo15RefinerLatent", + inputs=[ + io.Conditioning.Input("positive"), + io.Conditioning.Input("negative"), + io.Vae.Input("vae", optional=True), + io.Image.Input("start_image", optional=True), + io.ClipVisionOutput.Input("clip_vision_output", optional=True), + io.Latent.Input("latent"), + io.Float.Input("noise_augmentation", default=0.10, min=0.0, max=1.0, step=0.01), + + ], + outputs=[ + io.Conditioning.Output(display_name="positive"), + io.Conditioning.Output(display_name="negative"), + io.Latent.Output(display_name="latent"), + ], + ) + + @classmethod + def execute(cls, positive, negative, latent, noise_augmentation, vae=None, start_image=None, clip_vision_output=None) -> io.NodeOutput: + in_latent = latent["samples"] + in_channels = in_latent.shape[1] + cond_latent = torch.zeros([in_latent.shape[0], in_channels * 2 + 2, in_latent.shape[-3], in_latent.shape[-2], in_latent.shape[-1]], device=comfy.model_management.intermediate_device()) + cond_latent[:, in_channels + 1 : 2 * in_channels + 1] = in_latent + cond_latent[:, 2 * in_channels + 1] = 1 + if start_image is not None: + start_image = comfy.utils.common_upscale(start_image.movedim(-1, 1), in_latent.shape[-1] * 16, in_latent.shape[-2] * 16, "bilinear", "center").movedim(1, -1) + encoded = vae.encode(start_image[:, :, :, :3]) + cond_latent[:, :in_channels, :encoded.shape[2], :, :] = encoded + cond_latent[:, in_channels + 1, 0] = 1 + + positive = node_helpers.conditioning_set_values(positive, {"concat_latent_image": cond_latent, "noise_augmentation": noise_augmentation}) + negative = node_helpers.conditioning_set_values(negative, {"concat_latent_image": cond_latent, "noise_augmentation": noise_augmentation}) + if clip_vision_output is not None: + positive = node_helpers.conditioning_set_values(positive, {"clip_vision_output": clip_vision_output}) + negative = node_helpers.conditioning_set_values(negative, {"clip_vision_output": clip_vision_output}) + + return io.NodeOutput(positive, negative, latent) + + PROMPT_TEMPLATE_ENCODE_VIDEO_I2V = ( "<|start_header_id|>system<|end_header_id|>\n\n\nDescribe the video by detailing the following aspects according to the reference image: " "1. The main content and theme of the video." @@ -280,6 +324,7 @@ class HunyuanExtension(ComfyExtension): EmptyHunyuanLatentVideo, EmptyHunyuanVideo15Latent, HunyuanVideo15ImageToVideo, + HunyuanVideo15RefinerLatent, HunyuanImageToVideo, EmptyHunyuanImageLatent, HunyuanRefinerLatent,