MiniMax H3 in ComfyUI: PDD Acc LoRA (8-Step) vs Turbo (4-Step) on RTX 3090

Leer en español →

A small YouTube channel tested an 8-step acceleration LoRA Alibaba PAI released for MiniMax H3 — we confirmed it’s real: alibaba-pai/MiniMax-H3-Acc-LoRAs on HuggingFace, plus the custom node Jalen-Brunson/ComfyUI-MiniMax-H3-PDD-Acc you need to actually use it. The video’s central claim: the 4-step turbo LoRA we’ve used in earlier articles falls short in fast-action scenes.

We didn’t repeat its examples — we generated our own, with the same prompt and the same seed in every comparison, on our own RTX 3090. And the real finding isn’t the one we expected.

At a glance

The finding, in video

Same prompt (anime sword duel, rain-soaked neon rooftop), same seed 1007, same duration: 4-step turbo (left) vs. 8-step PDD Acc (right).

At first glance, any single frame from either version looks good — clean anime-style animation, good silhouette separation, energy effects consistent with the prompt. The difference doesn’t jump out in an isolated frame. It jumps out when you compare the SAME moment in time across both videos:

Top: 4-step turbo at second 7 and at second 12 of the same clip — the clash pose barely changes between the two moments. Bottom: 8-step PDD Acc at the same two moments — the characters have spatially separated, the action moved forward.

Across 5 seconds of clip, the 4-step version essentially keeps the same clash composition frozen — the red energy arc is still there, the distance between the characters barely changes. The 8-step version, over the same interval, shows real progression: spatial separation, motion continuity, the choreography advances. It’s not that the 4-step turbo produces ugly frames — it’s that it seems to “get stuck” on a pose instead of developing the action over time.

Confirmation with a second scene

Different fight scene, 5.17s, same prompt and seed across all three variants: 4-step turbo, 8-step PDD Acc, and the two-sampler hybrid trick.

With a completely different prompt and scene (a fight with clear foley instead of the sword duel), the pattern repeats: the mid-clip frame of the 4-step version shows a blurry, hard-to-read clinch, while PDD Acc and the hybrid both show a sharp, legible fighting stance at the same moment. Two scenes, two prompts, same pattern — not a coincidence from a single test.

Real installation: none of this is where you’d expect

Three things no regular LoRA tutorial prepares you for:

1. The folder isn’t models/loras/. It’s ComfyUI/models/pdd_acc/, and it probably doesn’t exist yet in your install — you’ll need to create it by hand with that exact name.
2. There are two files, not one. MiniMax-H3-FL2VA-Acc-8Step.safetensors and MiniMax-H3-Ref2VA-Acc-8Step.safetensors (1.37GB each). Using the one that doesn’t match your base checkpoint throws no error — the result just comes out wrong and looks like the LoRA is broken.
3. The custom node didn’t register until a full restart. We cloned Jalen-Brunson/ComfyUI-MiniMax-H3-PDD-Acc with no errors, but the MiniMaxH3PDDAccApply node wasn’t showing up as available. The cause: an earlier ComfyUI process had stayed alive, still listening on the same port after a normal kill — it took kill -9 and a clean relaunch for the node to actually register.

The node’s own README also demands a stricter sampler recipe than regular turbo: euler as the sampler, full stop (no er_sde, no dpmpp), sigmas have to come from the MiniMaxH3PDDAccApply node itself (not a generic BasicScheduler), the 1.0 CFG gets applied via BasicGuider instead of the KSampler’s cfg field, and it explicitly warns against stacking this LoRA with regular turbo LoRAs — they’re two different distillations that don’t combine well.

The hybrid trick: almost as good, but the timing number has a catch

The original video proposed chaining two samplers to save time without losing action quality: a first pass with the 8-step PDD Acc LoRA (drives the motion), then a second pass with the 4-step turbo at low denoise, just to refine detail. It didn’t give an exact denoise value — we used 0.3, our own call, not the original video’s.

Visually, the result is almost indistinguishable from pure PDD Acc — makes sense, it’s a light refinement pass on the same already-resolved action latent. But the timing figure (83.5s vs. pure PDD Acc’s 120.3s) needs a real caveat: our ComfyUI log shows “0 models unloaded” for that run, meaning it benefited from the model cache still warm from the PDD Acc run that ran immediately before it. This isn’t a clean cold comparison — we can’t confirm the savings would be as large starting from a freshly launched ComfyUI.

Time and VRAM: in our tests, 8-step consistently takes longer

The pattern repeats across two scenes of different length and content, on the same hardware and resolution (832×480): the 8-step LoRA takes between 51% and 60% longer than the 4-step turbo. Not a fluke from a single test. Real VRAM peak: 23.27GB with pure PDD Acc, 23.53GB with the hybrid — both leave little headroom on a 24GB card. We didn’t capture the 4-step baseline’s VRAM peak in the fight scene (the logger started late), so we can’t give a comparable VRAM figure for that variant.

Limitations of this test

• We couldn’t verify the original video’s claim about audio (that it sounds “unrealistic” with 4 steps) — we didn’t run a controlled listening evaluation.
• The hybrid trick’s timing was measured with a warm model cache, not cold — the real savings starting from a freshly launched ComfyUI could be smaller.
• The denoise value (0.3) for the hybrid’s second pass is ours, not the original video’s, which didn’t give an exact number.
• We only tested the hybrid trick on the fight scene, not on the sword duel — we don’t know if the “almost as good as pure PDD Acc” pattern holds on a clip three times longer.
• A single GPU (RTX 3090), no statistical repetition within each individual test — we did repeat across two different scenes and prompts, which is what gives us confidence in the pattern.
• We didn’t capture the 4-step baseline’s VRAM peak in the fight scene (an instrumentation miss, not a model issue).

Conclusion

Alibaba PAI isn’t overselling this entirely, but it’s not the simple “4-step is bad” story either. Across two different scenes and prompts, both LoRAs produce individually attractive frames — the real difference is whether the choreography advances over time or gets stuck on a pose. The 4-step turbo seems to freeze the action; the 8-step PDD Acc develops it, at the cost of 51-60% more generation time, consistently across both tests.

🏆 Our recommendation

If your scene depends on action progressing legibly over time (fights, complex movement, choreography), use the 8-step PDD Acc and accept the time cost. If your scene is more static or motion matters less than iteration speed, the 4-step turbo remains a reasonable choice — it doesn’t produce ugly frames, just less dynamic choreography. The hybrid trick is worth trying if you need that middle ground, but don’t trust the time savings we saw without measuring it yourself, cold. And before installing the LoRA, create the models/pdd_acc/ folder by hand and make sure you download the right file for your checkpoint (FL2VA or Ref2VA) — the wrong one gives no warning, just a worse result.

Keep Reading

If you haven’t tried MiniMax H3 on your RTX 3090 yet, start with our Sage Attention and Sol-Attn test. If you’re after swapping identities in video with this same model, we tested real face swap with Ref2VA, including the same kind of LoRA substitution we document here. And if you need videos longer than a few seconds, we tested Motion Director for chaining segments with continuity.