FastH3 OmniRef Accel is an unofficial ComfyUI node pack for a specific two-stage reference-video workflow. It retains eight sampling steps while changing the generation sizes, text-encoder execution, and decode/export path. Its purpose is to make a short portrait-oriented clip practical on the author's RTX 4070 setup, with visible quality tradeoffs for small faces.
The author's October 23 announcement links the original repository. This guide checks its validation record, workflow, and model references on that date. We have not reproduced its renders. It is separate from the official FastH3 release and from H3Vid's hosted generator.
Understand the speed comparison
The author's reference case is a 124-frame clip with audio and a 1024×1792 delivery frame. Their official-template comparison generates directly at that size and takes 459.6 seconds; the optimized pack reports about 64.6–66.8 seconds after server start. Server startup itself is excluded from the execution interval.
Those paths do different work. In the optimized recipe, the first five steps run at 256×448. An LBH latent upscaler then doubles the video latent, and the remaining three steps run at 512×896. H3's VAE decodes the result; Real-ESRGAN enlarges it before resize and crop produce the requested delivery dimensions.
The larger exported file therefore does not mean the diffusion model generated every step at full output resolution. The author explicitly reports that small faces in wide or full-body shots can lose detail that the direct template preserves. Start with a bust-up or knee-up subject when evaluating this recipe.
Match the documented environment
The validation machine uses Windows 11, an RTX 4070 with 12 GB VRAM, and 48 GB system RAM. The record pins a ComfyUI revision, Python, PyTorch, comfy-kitchen, and PyAV versions. Other GPUs, Linux, longer clips, and different ComfyUI versions remain outside that validation record.
Native H3 reference nodes and BlockSparseAttention must already work. The export path also needs PyAV with h264_nvenc. The documented launch keeps Dynamic VRAM enabled and uses --disable-pinned-memory. Treat these as a matched starting environment, not universal settings for every H3 installation.
The pack has no extra Python-package requirement, but it does require several existing model files. A successful node installation alone does not show that the encoder, latent upscaler, or export codec is available.
Select the complete model set
| Component | Documented file |
|---|---|
| Diffusion model | fasth3_omniref_8step_pruned_int8_convrot.safetensors |
| Text encoder | qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors |
| Video VAE | minimax_h3_video_vae_int8_convrot.safetensors |
| Audio VAE | minimax_h3_audio_vae_fp32.safetensors |
| Latent upscaler | minimax_h3_latent_upscaler_3d_fp16.safetensors |
| Pixel upscaler | realesr-animevideov3.pth |
The README links the diffusion, encoder, and VAEs to FastVideo, the latent model to LBH, and the final upscaler to Real-ESRGAN. Its validation file includes hashes. A newer LBH filename is accepted by the setup check, but the author does not establish that it is identical to the tested file.
The NVFP4 encoder name does not establish native FP4 acceleration on an RTX 4070. The author chose its smaller storage footprint partly to fit the working file cache alongside the diffusion checkpoint. Model download size, system-memory pressure, and GPU math support are separate considerations.
Check, install, and run a short reference case
Stop ComfyUI, then run the pack's setup.bat with explicit paths to the ComfyUI directory and its own Python interpreter. Add --check first for the read-only dependency check. Run the documented install command only after the expected nodes, models, and encoder path are identified.
Import workflows/fasth3_omniref_accel.json. Set the final dimensions in Output size, select the reference picture, and edit the Reference to Video prompt. Use the picture's numbered reference tag in the prompt. Start with one short shot and one image so you can diagnose identity, framing, and sound before adding complexity.
The sizing node enforces a minimum generation short side. Consequently, requesting 720×1280 does not simply divide both sides by two: the documented path generates at 512×896, resizes, and center-crops. Check the final framing if a face or prop sits near an edge.
Judge the tradeoff on the intended shot
Save the exact model hashes, prompt, reference, seed, and sizes with the MP4. Compare complete motion and sound, especially faces that become small during a camera move. The same seed across the direct and two-stage recipes can produce different composition and movement, so it is not a guarantee of pixel-matched scenes.
The author's bit-identical checks concern selected implementation changes, not equivalence of the entire optimized recipe to direct full-resolution generation. Keep that distinction when reporting results. For the official checkpoint and broader hardware choices, use the OmniRef setup guide; for the earlier latent handoff concept, see LBH 3+1.