Imagineer13/H3-Anti-Burn-LoRA

H3 Anti-Burn LoRA (universal, v1)

When you make a long MiniMax H3 video by chaining clips, each new clip continues from the end of the previous one. It copies that ending's small flaws and adds a few more. After a few clips, faces "burn": skin and lines get crunchy and etched, and the picture gets harsher. This LoRA is used on the continuation clips and keeps them close to the look of the first clip.

It was trained on three turbos (Larry Turbo v4, HyperFlow, lightx2v 8-step) and tested on a fourth model it never saw, where it cut face burn from +61% to +4%.

How to use it

• Load it with LoraLoaderModelOnly on the model that makes your continuation clips: 2, 3, 4 and so on.
• Don't use it on clip 1. A first clip has nothing to burn from. The LoRA was never trained or tested there.
• Start at 0.5. If faces still get crunchier from clip to clip, go up; turbos that burn harder need more. If faces turn soft or waxy, go down; turbos that burn less need less.
• How strongly it acts also depends on your setup. For example, merging it into an 8-bit checkpoint weakens it. So try a few values on a 4-5 clip chain before you settle.

Examples

Each image shows the same chain run without and with the LoRA. The prompts and seeds are the same, so clip 1 is identical; the LoRA is never used on it. Each frame is taken 1 second before the end of its video, about 37 seconds in. We picked scenes where the camera stays on the same shot, so the two clip-10 frames compare fairly.

Face burn in this scene: +51% without the LoRA, −4% with it at 0.6.

Face burn in this scene: +52% without the LoRA, +6% with it at 0.5.

Larry, the hardest burner we tested. Face burn in this scene: +189% without the LoRA, −15% with it at 0.6, loaded at full strength (not merged). With a turbo that burns this hard, the LoRA still helps a lot, but it can soften the image, and here the colour still drifted warmer and darker. Baked into an 8-bit checkpoint, 0.6 is not enough for Larry; go higher there.

What we measured

We ran 10-clip chains, each about 37 seconds long. There were 4 scenes per setting (kitchen, kitchen with a second seed, park, porch), and every number is the average of those chains. "Face burn" is how much the face's fine texture changed from clip 1 to clip 10:

• 0 means clip 10 looks as clean as clip 1.
• Positive means it burned.
• Negative means it went softer than clip 1.

Turbo	No LoRA	0.4	0.5	0.6	0.8
An unseen model (never trained on)	+61%		+29%	+4%	−13%
HyperFlow 8-step (LoRA)	+25%	+4%	−1%	−14%
lightx2v 8-step v1.0 768p	+44%	+14%	0%	−7%	−37%¹
Larry Turbo v4 (step 600 EMA)	+261%	+212%²		+159%²	+57%¹

¹ 2 chains. ² 3 chains: the park face got too small to measure.

Larry burns far harder than the others and needs more than 0.8 in this setup. At 1.0 its face burn was gone (−19%, 2 chains), but the camera framing started to change. Loaded at full strength (not merged) instead, 0.6 took it from +261% to −5% (3 chains).

What else changes at the best strength. Saturation, skin redness and lip sync stay where they are without the LoRA. Fine detail drops a little: 4–14% below clip 1, against 3–18% above it without the LoRA. Too much strength shows up as soft, waxy faces.

The setup we tested.

• Checkpoints: ComfyUI's int8 ConvRot pruned MiniMax H3 checkpoints (Comfy-Org/MiniMax-H3). Clip 1 used fl2va; continuations used ref2va.
• LoRA: merged with LoraLoaderModelOnly before the turbo.
• Continuations: each one carried the previous clip's last 39 frames and their audio as a protected masked context.
• Sampling: 8 steps, cfg 1.
• Resolution: 640×1152, 1152×640 and 864×864.

With other checkpoints (bf16, fp8, GGUF), other step counts or other ways of continuing, expect a different best strength.

Limitations

• It is experimental. Four scenes per setting is a small test, and one 10-clip chain can swing ±15 points on its own.
• It prevents burn. It does not repair a video that is already burned.
• It was only tested on continuations from masked context, not on extending a video any other way.
• It was trained and tested on speaking people, with faces at medium to close distance.

How it was trained

It uses trajectory self-distillation:

1. We recorded what each turbo actually does in a fresh generation: every model call of real 8-step runs. That gave 150 full-size runs, 50 each from Larry, HyperFlow and lightx2v.
2. The LoRA was trained in continuation mode: the start of the same video is pinned as the protected context (sometimes slightly noised), and the target is the turbo's own fresh run. So a continuation learns to look like a fresh generation instead of copying and amplifying the flaws in its context.
3. There was no synthetic burn and no fresh-noise loss. Earlier versions that used them softened everything.

Run settings:

• 300 steps, rank 32, learning rate 1e-4, network weight 0.6.
• Resolution 864, 124 frames.
• One RTX PRO 6000 (96 GB).

The training code, configs, recorder node, scoring script and a write-up of everything that didn't work are in TryingSomeStuff13/ai-toolkit, branch h3-antiburn.

File

File	h3_antiburn_universal_v1.safetensors
Size	310,168,728 bytes
SHA256	8f76dc3c63afa8c556c620b4b7f2ffeeb666f0aa9382a9ce3a07fd6120d0f154
Type	LoRA, rank 32, bf16, 416 tensors, ComfyUI key format (diffusion_model.*)
Trained layers	attention and MLP of the 50 DiT blocks and the 2 token-refiner blocks; no adaLN

License

This LoRA is a Model Derivative of MiniMax H3 and is distributed under the MiniMax H3 Community License Agreement (see LICENSE and NOTICE), including its Acceptable Use Policy. The file holds only new adapter weights; no MiniMax H3 weights are included.

Thanks to a community compute grant that paid for the GPU time.