MiniMax H3 TaoMate LoRA, Singularity and turbo: which pairings hold colour

On MiniMax H3, the TaoMate 3-step LoRA held its colour on the stock pruned checkpoint (mean saturation 42.8, warmth −3.5) and lost it on the Singularity v1.3 checkpoint (28.6 and +13.8) with the same seed, prompt and reference image. Singularity with the lightx2v turbo LoRAs measured 46.2 to 47.2, so the finding is about one pairing, tested once per configuration on one RTX 5090.

Which LoRA and checkpoint pairings hold colour on MiniMax H3?

Every pairing but one held. Stock with turbo, stock with TaoMate and Singularity with turbo all stayed cool and saturated, which fits the scene: a rainy neon street at night. Singularity with TaoMate did not.

Six runs sit in a narrow band: saturation 42.8 to 47.4, warmth −3.5 to −12.2, brightness 87.1 to 94.9. Runs 2 and 8, the two that put TaoMate on Singularity, are the only ones outside it, with the only positive warmth values and the only brightness above 100. In those two clips, by eye, the character's red hair turns brown, the street hazes over, and a microphone appears in her hand that is in neither the prompt nor the reference image.

The 4-step and the 8-step turbo LoRAs on Singularity landed 1.0 saturation point and 3.3 warmth points apart. That is one clip each and two colour metrics, not a verdict on either LoRA.

Test setup

• MiniMax H3 reference-to-video in ComfyUI (MiniMaxH3ReferenceToVideo), 1280×720, 243 frames at 24 fps, ref_image_size set to match.
• Seed 20260918, fixed. One prompt, one reference image and one voice reference clip, identical in every run.
• SamplerCustomAdvanced with BasicGuider, BasicScheduler on simple, denoise 1.0, LoRA strength 1.0.
• Text encoder qwen3vl_32b_minimax_h3_int8_convrot.safetensors, VAEs minimax_h3_video_vae_fp16.safetensors and minimax_h3_audio_vae_fp32.safetensors.
• Runs 6 to 8 add a MiniMaxH3SigmaShift node (shift_video 12.0, shift_audio 3.0) and use euler.

Every setting was read back out of the workflow ComfyUI embeds in each MP4, not from notes.

Colour comes from a small Python script that decodes each clip through ffmpeg at 2 frames per second, scaled to 96×54, and averages the 8-bit red, green and blue values. Brightness is the mean of the three, warmth is mean red minus mean blue, and saturation is the mean HSV saturation of every 37th pixel of those frames, times 100. These averages say nothing about motion, identity or lip sync.

Tested 2026-09-18 on an RTX 5090 (32 GB), 96 GB RAM, Windows 11, ComfyUI 0.36.0. One seed, one prompt and one reference image, one run per setting.

Is TaoMate the problem, or Singularity?

Neither on its own. The pairing is the problem, and two comparisons isolate it.

• Same LoRA, different checkpoint (runs 5 and 2). TaoMate on stock: 42.8 saturation, −3.5 warmth. TaoMate on Singularity: 28.6 and +13.8. That is about a third of the saturation gone with only the checkpoint changed.
• Same checkpoint, different LoRA (runs 4 and 2). Singularity with turbo 4-step: 47.2 and −7.4. Singularity with TaoMate: 28.6 and +13.8. The step count moves from 4 to 3 here because each LoRA runs at the steps it was distilled for.

Each part is fine with another partner. The upstream pages give some context.

TaoMate-H3, from the Alibaba TaoLive AIGC team, is a low-latency streaming runtime built on MiniMax H3 with a 3-step LoRA. That LoRA ran at the lowest step count in this test and held colour on the stock checkpoint. The released weights are the text-to-video adapter, the README loads it on the base FL2VA checkpoint, and the reference-to-video version is listed as TBD, so my reference-to-video runs are already outside what it was released for. The file I used is the ComfyUI conversion published by Kijai. Its name says average rank 19, while the upstream adapter is rank 128. I have not tested the full-rank adapter.

Singularity, by WarmBloodAban, is described on its card as a fusion of several H3 checkpoints that was then fine-tuned and pruned, aimed at cleaner image quality, distant faces and fast motion. Its weights are no longer the base weights. The card recommends minimax_h3_ref2v_turbo_4step_v0.1 as its speed LoRA, which is the pairing that measured 47.2 here.

My reading, which this test does not prove: a step-distillation LoRA is a shortcut through the sampling path of the weights it was trained on. Move the weights underneath it and the shortcut can land somewhere else. Not always, though. The lightx2v turbo LoRAs are distilled from the base model too, and they held on Singularity. A changed base is a reason to check a pairing, not a guarantee that it loses colour.

Does changing the sampler or sigma shift fix it?

Not in the one run I have, and this round is an observation, not a controlled result. For runs 6 to 8 I changed two things at once: the sampler went from res_multistep to euler, and I added the sigma shift node at 12.0 for video and 3.0 for audio, the values lightx2v lists as the training shifts for the Ref2VA Turbo 4-step v0.1. I cannot say which of the two caused any difference.

With that caveat, the pairings that were already holding moved by less than 3 saturation points: Singularity with turbo 4-step went from 47.2 to 46.7, stock with TaoMate from 42.8 to 45.5. Singularity with TaoMate stayed warm and washed out: 28.6 to 25.5 saturation, +13.8 to +11.2 warmth. Sampler and scheduler choices have their own post: MiniMax H3 sampler and scheduler settings.

Does the pairing change the audio timing?

In this test, yes: the timing follows the checkpoint and LoRA pair, not the sampler round. Singularity with TaoMate stands out, with the earliest end to the first pause and the most total silence. H3 generates the audio in the same pass as the video, so I measured that too, with ffmpeg's silence detector at −30 dB and a 0.3 s minimum gap:

ffmpeg -vn -i clip.mp4 -af silencedetect=noise=-30dB:d=0.3 -f null -

The video calls the fourth column speech onset. Strictly it is the moment the first detected pause ends. Every track has sound above −30 dB before that pause, and the detector cannot tell speech from ambience.

Each of the three pairs I ran in both rounds repeats its first pause end within 0.01 s. An early marker does not mean less dead air either. Singularity with TaoMate gets there first, at 1.19 s, and carries the most total silence, 2.43 s and 2.83 s. Stock with turbo 4-step has the least, 1.20 s.

What I would run

• On Singularity v1.3: the ref2v turbo 4-step v0.1 LoRA at 4 steps. It is the pairing the Singularity card recommends, and it measured within 0.2 saturation points of the stock checkpoint with the same LoRA.
• With TaoMate 3-step: the stock pruned ref2va checkpoint, until a pairing with a finetune has been checked.
• Before any new pairing goes into a batch: render one clip on a fixed seed next to a pairing you trust and compare the colour first. In this test the drift showed up plainly in average saturation and warmth.

For a timed comparison of two other speed LoRAs, see the FastH3 versus turbo LoRA test. LoRAs and LoRA strength are covered in Part 2 of ComfyUI From Zero.

Sources and files

• TaoMate-H3 by the Alibaba TaoLive AIGC team: the streaming runtime, its 3-step LoRA and the release table.
• Minimax-h3_Singularity by WarmBloodAban, on Hugging Face as WarmBloodAban/Minimax-h3_Singularity: the model card and the v1.3 pruned int8 file.
• Kijai's ComfyUI conversions, on Hugging Face as Kijai/MiniMax-H3_comfy: the TaoMate 3-step LoRA file used here.
• Minimax-H3-Turbo (ModelTC on GitHub, lightx2v on Hugging Face): model specs for the turbo LoRAs, including training shifts.
• MiniMax H3: the base model card.
• ComfyUI MiniMax H3 tutorial: points to the Comfy-Org/MiniMax-H3 repository on Hugging Face, where the stock pruned int8 checkpoint is published.
• The video this post belongs to, with the clips and their generated audio.
• Stubelius Ultimate H3: my free H3 workflow pack, where the speed LoRA is set on the Models node. The test itself used a plain ComfyUI graph of core nodes.
• Related posts: FastH3 vs turbo LoRA on MiniMax H3 and MiniMax H3 sampler and scheduler settings.

Common questions