## Summary
Sampling a long SeedVR2 video in one pass runs out of VRAM: the DiT working set grows linearly with latent frames times pixel area, so a 100 frame clip at 4x upscale needs more memory than any consumer card has. This PR adds two workflow-level nodes that split the latent into overlapping temporal chunks and recombine the sampled chunks with a Hann crossfade. The executor's list handling runs the stock KSampler once per chunk, so the sampler itself is untouched.
- **Chunk SeedVR2 Latent** splits the latent on the temporal axis. `frames_per_chunk` is in pixel frames on the 4n+1 grid, `temporal_overlap` sets how many latent frames adjacent chunks share, and `chunking_mode=auto` solves the chunk size from measured free VRAM and the latent's own dimensions. The node outputs the effective overlap so the merge is wired, not typed.
- **Merge SeedVR2 Latent Chunks** recombines the sampled chunks in order, crossfading each shared region with a Hann window (flat shoulders on the outer thirds, fade across the middle third). Zero overlap is a plain concatenation, bit-identical to `torch.cat`.
## Changes
- Added 2 nodes and 1 crossfade helper to `comfy_extras/nodes_seedvr.py` (+201 lines).
- Added 4 chunk-law constants to `comfy/ldm/seedvr/constants.py` (+10 lines).
- Added pytest unit tests in `tests-unit/comfy_extras_test/test_seedvr2_temporal_chunk.py` (+66 lines): chunk geometry, 4n+1 and mode validation, overlap clamping, temporal noise mask slicing (5-D and 4-D), the auto law including batch scaling, crossfade weights and blend direction, round trip, and metadata handling.
Everything outside the two new nodes is byte-identical to the base branch. The new constants are read only by the chunk node, and workflows that do not use these nodes take no new code path.
## Auto mode calibration
The auto law is `max_latent_frames = (free_GiB - reserved - margin) / (0.30 * megapixels)`, calibrated on an RTX 5090 (32 GB) with the 3b fp16 model: a 17-cell resolution sweep plus temporal bisection located the activation wall and confirmed it is linear in latent frames times pixel area (the same total-voxel budget holds from 1.5:1 through 24:1 aspect ratios and under transposition). The margin is four standard deviations of the measured run-to-run spread, which costs about one latent frame of chunk size and makes an out-of-memory failure a lottery ticket rather than a coin flip. Manual mode bypasses the law entirely.
On a 32 GB card, a 640x480x100 input at 4x upscale sampled as a single chunk allocates past 31 GiB and dies; auto mode picks 49 frame chunks (three chunks with overlap) and the full pipeline completes in about 240 seconds. The same law stands down on small inputs: a 320x240x100 clip runs as a single chunk because it fits.
## Example workflow
Load a video, 4x upscale, auto chunking, temporal overlap 3. Expected output for a 640x480x100 input: 2560x1920, 100 frames, seams invisible at the default overlap.
<details>
<summary>API workflow JSON</summary>
```json
{
"14": {"inputs": {"vae_name": "ema_vae_fp16.safetensors"}, "class_type": "VAELoader"},
"17": {"inputs": {"video": ["24", 0]}, "class_type": "GetVideoComponents"},
"20": {"inputs": {"fps": ["17", 2], "images": ["30", 0], "audio": ["17", 1]}, "class_type": "CreateVideo"},
"21": {"inputs": {"tile_size": 192, "overlap": 64, "temporal_size": 64, "temporal_overlap": 8, "pixels": ["27", 0], "vae": ["14", 0]}, "class_type": "VAEEncodeTiled"},
"22": {"inputs": {"tile_size": 256, "overlap": 32, "temporal_size": 64, "temporal_overlap": 8, "samples": ["33", 0], "vae": ["14", 0]}, "class_type": "VAEDecodeTiled"},
"23": {"inputs": {"unet_name": "seedvr2_3b_fp16.safetensors", "weight_dtype": "default"}, "class_type": "UNETLoader"},
"24": {"inputs": {"file": "input.mp4", "video-preview": ""}, "class_type": "LoadVideo"},
"25": {"inputs": {"filename_prefix": "video/seedvr2_upscale", "format": "auto", "codec": "auto", "video": ["20", 0]}, "class_type": "SaveVideo"},
"26": {"inputs": {"resize_type": "scale by multiplier", "resize_type.multiplier": 4, "scale_method": "bicubic", "input": ["17", 0]}, "class_type": "ResizeImageMaskNode"},
"27": {"inputs": {"resized_images": ["26", 0]}, "class_type": "SeedVR2Preprocess"},
"28": {"inputs": {"model": ["23", 0], "vae_conditioning": ["32", 0]}, "class_type": "SeedVR2Conditioning"},
"29": {"inputs": {"seed": 5770521, "steps": 1, "cfg": 1, "sampler_name": "euler", "scheduler": "simple", "denoise": 1, "model": ["23", 0], "positive": ["28", 0], "negative": ["28", 1], "latent_image": ["32", 0]}, "class_type": "KSampler"},
"30": {"inputs": {"color_correction_method": "lab", "images": ["22", 0], "original_resized_images": ["26", 0]}, "class_type": "SeedVR2PostProcessing"},
"32": {"inputs": {"frames_per_chunk": 21, "temporal_overlap": 3, "chunking_mode": "auto", "latent": ["21", 0]}, "class_type": "SeedVR2TemporalChunk"},
"33": {"inputs": {"temporal_overlap": ["32", 1], "latent_chunks": ["29", 0]}, "class_type": "SeedVR2TemporalMerge"}
}
```
</details>
## Prior art
- Reference implementation: https://github.com/ByteDance-Seed/SeedVR
- Community precedent for temporal chunking with blended reassembly: https://github.com/numz/ComfyUI-SeedVR2_VideoUpscaler
Chunk boundaries are a mathematical compromise: the model attends within a chunk, so different chunkings produce different outputs. The overlap crossfade hides the seam; power users can widen or zero the overlap from the workflow.
Add a --enable-asset-hashing CLI flag (action=store_true, default False)
and plumb it into the two asset-seeder call sites in main.py that
previously hardcoded compute_hashes=True (the startup scan and the
post-job output enqueue). Local runs now skip blake3 hashing unless the
user opts in, avoiding the startup/per-output cost on large models
directories while keeping hashing available for asset-portability
features.
Co-authored-by: Alexis Rolland <alexisrolland@hotmail.com>
* main: implement --vram-headroom
Implement --vram-headroom for dynamic vram as a hybrid debug/diagnostic
option that can be used for people who still report shared VRAM spills.
They can trial and error the setting to maintain a bit more headroom to
avoid shared VRAM spills.
* main: implement --reserve-vram
Implement --reserve-vram as extra headroom on the simple method which
is semantically as close as possible to the stated functionality and
formet behaviour of non-dynamic VRAM.
Add this option for users who know they have so much ram they want
to pin everything or have a pagefile that outruns their disk speed.
The removes the RAM pressure caps completely and pins behind the
primary model load forcing all models to be permanently comitted
to RAM.
Some custom nodes .to weights completely out of load context which
can wreak havoc if its for a model that is not active. Detect this
condition and just let it fall-through to the non-dynamic loader
straight up.
Some custom nodes try to set this true globally. It messes with dynamic
VRAM with one-off spikes that can OOM but this is also very high risk
for windows where such allocations might get serviced by shared memory
fallback.
Trump it.
cleanup_models_gc can be called once per load_models_gpu via
free_memory, which in turn can de-activate an active model via
this reset_cast_buffers.
cleanup_models_gc() could also come via obscure garbage collector
paths so limit reset_cast_buffers to the post-node callsite instead.
* mm: split off registration helper to doer and headroom calc
* pinned_memory: implement registration comfy side
Move away from Aimdo buffer registrations which seem fraught with
danger and do it comfy side. Just start with the basic move.
* pinned_memory: do registrations as portable memory
* pinned_memory: discard async errors on registration fail
Like the good ol days.
* pinned_memory: implement abs shortfall retry
If pinned registration happens to fail despite the previous budget
ensures, consider the allocation shortfall, ensure it again, and
try again. This allows comfy pins to interoperate with other software
that might be doing substantive pinning.
