☕ One coffee = one happy creator: https://buymeacoffee.com/the_frizzy1

Your video, a light sphere and a line of light text go in, the same shot comes back in a new light. Same people, same room, same camera, same timing, original sound. It is a local ComfyUI rebuild of Sirio Berati's open-source relight, which was built for one rented 80 GB GPU, on MiniMax H3 Ref2VA int8 with the ref2v turbo LoRA.

What you need

✔ Custom nodes — None. Core ComfyUI only.
✔ ComfyUI version — 0.44 or newer. 0.44.0 for the renders, 0.45.0 with frontend 1.58.5 for the saved file
✔ VRAM — 12 GB — tested, not estimated. Peak 11,841–11,945 MiB
✔ System RAM — 32 GB, 21.4–22.6 GB peak in the runs below
✔ Tested on — RTX 3060 12 GB · 32 GB DDR4 · ComfyUI 0.44.0 (renders) · ComfyUI 0.45.0 + frontend 1.58.5 (the saved file)
✔ Licence — Workflow MIT · light presets and spheres MIT, Sirio Berati · MiniMax H3 and the NVIDIA latent upscaler: MiniMax H3 Community License · Qwen3-VL Apache-2.0 · Real-ESRGAN BSD-3-Clause

Models to download

Click a file name to download it. Put it in the folder shown, inside ComfyUI/models/. Every file name was read from the workflow .json. If you already run H3 and an H3 swap workflow, you have most of them.

Required — the first six files, 51.5 GB. That is everything with Full HD off.

diffusion_models/

File: minimax_h3_ref2va_pruned_int8_convrot.safetensors — 21.0 GB
Download:
https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/diffusion_models/minimax_h3_ref2va_pruned_int8_convrot.safetensors

Put it here:
ComfyUI/models/diffusion_models/

loras/

File: minimax_h3_ref2v_turbo_4step_v0.1_comfyui_bf16.safetensors — 2.0 GB
Download:
https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/loras/minimax_h3_ref2v_turbo_4step_v0.1_comfyui_bf16.safetensors

Put it here:
ComfyUI/models/loras/

text_encoders/

File: qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors — 15.7 GB
Download:
https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/text_encoders/qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors

Put it here:
ComfyUI/models/text_encoders/

File: qwen3vl_8b_int8_convrot.safetensors — 9.4 GB (all (the describer))
Download:
https://huggingface.co/Comfy-Org/Qwen3-VL/resolve/main/text_encoders/qwen3vl_8b_int8_convrot.safetensors

Put it here:
ComfyUI/models/text_encoders/

vae/

File: minimax_h3_video_vae_int8_convrot.safetensors — 2.8 GB
Download:
https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/vae/minimax_h3_video_vae_int8_convrot.safetensors

Put it here:
ComfyUI/models/vae/

File: minimax_h3_audio_vae_fp32.safetensors — 0.6 GB
Download:
https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/vae/minimax_h3_audio_vae_fp32.safetensors

Put it here:
ComfyUI/models/vae/

Full HD — optional. The gentle refine adds 23.7 GB, RealESRGAN adds 0.1 GB.

diffusion_models/

File: minimax_h3_fl2va_pruned_int8_convrot.safetensors — 21.0 GB (Full HD: gentle refine)
Download:
https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/diffusion_models/minimax_h3_fl2va_pruned_int8_convrot.safetensors

Put it here:
ComfyUI/models/diffusion_models/

latent_upscale_models/

File: minimax_h3_latent_upscaler_3d_conv_v1_bf16.safetensors — 0.7 GB (Full HD: gentle refine)
Download:
https://huggingface.co/Efficient-Large-Model/Open-H3-Regenerate-2K/resolve/main/latent_upsampler/minimax_h3_latent_upscaler_3d_conv_v1_bf16.safetensors

Put it here:
ComfyUI/models/latent_upscale_models/

loras/

File: minimax_h3_open_regenerate_2k_r128_comfyui_bf16_swiglu_fixed.safetensors — 2.0 GB (Full HD: gentle refine)
Download:
https://huggingface.co/DrakenZA/Open-H3-Regenerate-2K-ComfyUI/resolve/main/minimax_h3_open_regenerate_2k_r128_comfyui_bf16_swiglu_fixed.safetensors

Put it here:
ComfyUI/models/loras/

upscale_models/

File: RealESRGAN_x2plus.pth — 0.1 GB (Full HD: RealESRGAN x2)
Download:
https://github.com/xinntao/Real-ESRGAN/releases/download/v0.2.1/RealESRGAN_x2plus.pth

Put it here:
ComfyUI/models/upscale_models/

The light presets and spheres are not models. Get them from sirioberati/relight-open-source (MIT): a preset's sphere image goes in the Light sphere input, its master text from presets.json in the Light text input.

Settings

You see five nodes on load: your three inputs on the left, the H3 Relight node with every setting, the Result on the right. Your video is read as 24 fps, up to 15 s in one pass. The video lands in ComfyUI/output/H3 relight/.

Short side 480 (fast) or 736 (sharper) · Seconds 0 = the whole clip, up to 15 s or 362 frames in one pass · Steps 6 with the turbo LoRA · Seed 42, the same seed for every pass of a long clip · Extra direction one more line for the prompt writer · Full HD Off, Gentle H3 refine (sigma 0.55) or RealESRGAN x2 · Long clip off, on = up to 60 s in 5.2 s passes. All 10 model loaders are on the same node.

Full HD means short side 1088, so 1920x1088 for a 16:9 clip. The gentle refine runs the H3 latent upscaler to 1088p then 3 Regenerate-2K steps on H3 fl2va from sigma 0.55 at CFG 1, the NVIDIA Open-H3-Regenerate-2K recipe at 1080p instead of 1440p. RealESRGAN x2 upscales every draft frame and Lanczos to 1088p, redrawing nothing. Only the path you pick runs.

Long clip relights in passes of 124 frames with core Start Loop / End Loop: the prompt is written once from the first frame, from pass 2 on the last 22 frames of the pass before are anchored at the start and then dropped, and each pass is stored with its own slice of the original sound before the passes are joined in order.

Speed on my RTX 3060

5.2 s desk clip (1920x1088), 6 steps, Blue Amber Bed preset, cold start. VRAM is the whole GPU, RAM the ComfyUI process.

• 480, Full HD off (3 presets) — 6:40 – 6:49 wall, 11,841 MiB peak vram, 21.4 GB peak ram

• 736, Full HD off (3 presets) — 21:03 – 24:02 wall, 11,938 MiB peak vram, 22.6 GB peak ram

• 736 + gentle refine to 1920x1088 — 31:48 wall, 11,945 MiB peak vram, 22.5 GB peak ram

• RealESRGAN x2 of a finished 736 clip (its own graph) — 3:54 – 5:49 wall, 7,744 MiB peak vram, 14.4 GB peak ram

736 is sharper, it is also three times the wait. The gentle refine to Full HD puts that at 31:48 for one 5.2 s clip.

What did not work

• NVIDIA's default start sigma of 0.909 redrew the face in my test, so the refine starts at 0.55 instead. The lower start keeps more of the draft.

• Long clip has not been rendered on the GPU yet, so there are no timings for it. A 60 s clip is 14 passes, so about 14 times one 480 run. Full HD does not apply to long clips.

• RealESRGAN inside this workflow has not been rendered either, only in its own graph, the row above.

• Clips are read as 24 fps. One pass is 15 s at most, 362 frames; longer needs Long clip. Seconds must not be longer than your clip.

• Frontend 1.58.5 put 14 links on the wrong input when it built the subgraphs, on Math Expression, MiniMax H3 Reference to Video and Start Loop. They were moved back before saving with source/relight_fix.js. The saved file was then loaded in 0.45.0, turned into a run graph with the frontend's own graphToPrompt for four settings and passed ComfyUI's validate_prompt for all four; 480 off and 736 gentle were compared node by node with the tested graph and come out as the same computation.

Get the workflow

• GitHub: https://github.com/Thefrizzy1/The_frizzy1-Workflows/tree/main/workflows/minimax/h3-relight-12gb

• Hugging Face mirror: https://huggingface.co/The-frizzy1/The_frizzy1-Workflows/tree/main/workflows/minimax/h3-relight-12gb

More videos on this stuff: https://www.youtube.com/playlist?list=UUGKDBMhTGqvlBZE9cF5d50Q