Maintained LoRA acceleration effectiveness while ensuring high-quality testing; this has been completed
After extensive testing and experimentation aimed at achieving higher-quality outputs with Minimax H3 when using Turbo LoRA, I have tried numerous approaches. Ultimately, I found that a dual‑sampler, two‑stage sampling workflow gives me significantly more control over the final results.
Here is the setup that worked best for me:
Stage 1 – First Sampler (Rough Fitting)
Checkpoint: minimax_h3_turbo_4step_ckpt850.safetensors
Weight: 1.0
Steps: 5–7
In this initial pass, the model rapidly fits the overall layout, content, and global scene logic. From my preliminary observations, the high‑variance sigma in this stage greatly influences scene composition, physical effects, and motion/dynamic understanding — which in turn directly affects artifact generation.
Therefore, I do not recommend using minimax_h3_turbo_4step_ema_ckpt850.safetensors for this first step.
Stage 2 – Second Sampler (Refinement)
Checkpoint: minimax_h3_turbo_4step_ckpt500.safetensors
(This version currently offers a good balance: it maintains acceptable dynamic behaviour while providing some sharpening and quality restoration.)
Weight: 0.7
(I found that values above 0.7 trigger noticeable denoising effects, while values below 0.69 make denoising almost ineffective; unfortunately, intermediate values do not seem to work well.)
Steps: 7–8
(My tests show that 7–8 steps are the sweet spot for this second stage.)
This second pass refines the details and further interprets high‑motion blur, building upon the foundation laid in the first stage.
Results & Observations
With this two‑stage workflow, I was able to:
Maintain quality without obvious loss in the model's physical understanding.
Avoid artifacts entirely.
Keep the dynamic/motion fitting close to the optimal threshold.
Retain the significant speed‑up provided by Turbo LoRA.
This approach is also an attempt to apply a training‑oriented mindset to the generation process — and it seems to work remarkably well.
I hope this sharing is helpful for the community, especially for those working on Minimax H3 Turbo LoRA and Minimax H3 extensions. If you find it useful and decide to adopt it, please feel free to cite me as @ANe5s .
Happy generating!（Translated via Deepseek）
@ANe5s Nice, But something doesn't seem right to me. The whole point of a 4-step LoRA is to reduce the number of steps, yet your two sampling passes take up around 15 steps—which defeats the purpose of the speed-up.
@ANe5s Nice, But something doesn't seem right to me. The whole point of a 4-step LoRA is to reduce the number of steps, yet your two sampling passes take up around 15 steps—which defeats the purpose of the speed-up.
Additional thoughts on the dual‑sampler workflow and the current checkpoint quirks
In my view, the core purpose of an acceleration LoRA is to reduce the number of fitting steps while maintaining or even improving output quality. At this stage, much of what I’m doing is exploring new ideas and uncovering behaviours in this still‑unfinished Turbo implementation.
Through my tests, I’ve observed the following:
ckpt850_emasuffers from ghosting/afterimage artifacts when used in a single‑sampler setup. To resolve that, you need to increase the step count to ~15.ckpt500(non‑ema), when used alone, also requires around 14–15 steps to avoid insufficient fitting during motion blur.
So, in practice, this two‑stage workflow improves efficiency by roughly 2–3 steps under the same conditions.
If your target sweet spot is 4–8 steps in total, you can simply adjust the step counts in each sampler and also tweak the sigma separation per stage. Moreover, in the second‑sampler LoRA, you have the flexibility to choose the sharpening intensity and overall sharpening effect you desire.
Current checkpoint behaviours in terms of sigma separation:
- With
ckpt500, the issue arises when sigma is low‑variance — it requires extra steps to compensate. - With
ckpt850_ema, the problem occurs when sigma is high‑variance; low‑variance cases also have minor issues that need a few more steps.
These findings are part of my ongoing experimentation, and I hope they help clarify the trade‑offs and possibilities for those working with Minimax H3 Turbo LoRA.
Feel free to use or adapt these insights — and if they prove useful, a citation to @ANe5s would be appreciated.
Total steps：13step
First sampling：minimax_h3_trubo_4step_ckpt500 strength：1.00 5step
Second sampling：minimax_h3_trubo_4step_ckpt500 strength：0.65 8step
672*1216，24 frames，15s
Prompt source： https://x.com/noorlewisx/status/2083081579262591469