joeygambino

/

MiniMax-H3-Multishot-Workflow

like 64

You're pushing the LoRA in every shot as the memory records the last frame, the one most impacted by the LoRA.

To keep things proper in the video, remember to use the proper label sections:

1. Integrated_multimodal_description:
2. overall_soundscape:
3. non_diegetic_music: N/A

I've added 4 step lora, but then result is a messed audio noise, as no way to add proper MiniMax-H3 Turbo Sampler (4-step) ... Any way to make it work with that lora, as that speeds up process a lot

PS: I did tests and probably Lora may not work well as it is. At each step image get more and more oversharpen, so first 10 sec is OK, then 10-20s Lora again sharpen last frame input I guess and gets over-sharpen and at 20-30s it is very over-over-sharpen ... Which I suspect is Lora fault and may not be fixable in such process.

The progressive oversaturation is the seam sharpening ratchet, and it is not specific to your LoRA: every content-continuous seam steps up in texture energy, because each shot's tail becomes the next shot's anchor and the model generates its continuation slightly sharper than what it was given. Gain above 1 in an autoregressive loop compounds - measured here at roughly 1.2–1.3x per join, which is invisible at 3 shots and obvious by 8. A LoRA that adds detail rides that loop and makes it arrive sooner.

Two levers for it, both already in the release:

• chain_gain_control = flatten - measures the texture level of shot 1 and holds every later shot to it. This is the one to reach for first.
• bank_ref_noise - a little seeded noise on the bank's reference clips, so they read as camera capture rather than as a pristine target to be enhanced.
  The noised clip never reaches the final cut. Try ~0.02–0.05 if flatten alone does not settle it.

The 4-step sampler part is a real gap and you are the second person to hit it - see issue #14 on the GitHub repo. The samplers built the sigma schedule themselves from the steps and scheduler widgets, with no way to supply your own, so a turbo LoRA that ships the schedule it needs could not use it. The next release adds a sigmas input to both multishot samplers: connect your own schedule and it replaces steps/scheduler entirely, and the console says so rather than silently overriding you.

On the audio noise specifically - my suspicion, and it is only that until you can test it, is that this is the same schedule problem rather than a separate one. A 4-step LoRA driven by a schedule shaped for 14 steps under-denoises, and the audio latent is far smaller and less redundant than the video one, so it shows the shortfall first and as noise. The sigmas input is what will let you find out. If it still comes out noisy on the LoRA's own schedule, please say so and post the schedule - that would point at the LoRA's audio branch instead, which is a different fix.