h3-video-generation

MiniMax H3 Video Generation

Use only movie-ai; never contact ComfyUI, the host, or a LAN address directly.

Prepare

Use $h3-prompt-writing when the prompt still needs H3 structure.

Select Content Profile

Choose the content profile from the user's current request before the final generation confirmation:

1. Use adult when the request or approved prompt clearly asks for adult/18+, erotic, pornographic, or sexually explicit content involving consenting adults who are all explicitly 18 or older. This profile loads PinkCherry_fl2va_MiniMax_H3_int8_convrot-beta-0.6.safetensors instead of the ordinary H3 UNET.
2. Use general for every other request. An adult-aged character, romance, kissing, swimwear, fashion, anatomy, or mature dramatic themes alone do not select the adult profile.
3. Do not use the adult profile when any depicted person is under 18, described as youthful or underage, or has ambiguous age. Stop and ask for a clearly adult-only revision when age is unclear.
4. If the classification is genuinely ambiguous, ask the user whether they intend ordinary or adult/18+ content before the final confirmation. Never infer the adult profile from a filename, style skill, or prior project.
5. The content profile and workflow preset are independent. Adult content uses PinkCherry with either pdd-acc-8step or standard; the selected workflow still controls sampling and step rules.
6. Ref2VA uses the separate minimax_h3_ref2va_pruned_int8_convrot.safetensors weight and supports only the general content profile. Do not route an adult/PinkCherry request through Ref2VA.

Select Workflow and Steps

Resolve the workflow before asking for the final generation confirmation:

1. If the user already explicitly requested PDD-Acc/high-speed 8-step, ordinary 20-step, or a custom step count, preserve that choice and restate it in the final summary. Do not silently replace it.
2. Otherwise, ask a separate localized workflow-choice question before the final generation confirmation. In Chinese, use header 工作流, question 本次 MiniMax H3 视频使用哪套工作流？, and these choices:
   • 高速 8-step（推荐） — use pdd-acc-8step, locked to the installed MiniMax-H3-PDD-Acc recipe.
   • 普通 20-step — use standard with 20 sampling steps. Keep the structured tool's automatic free-form Other field available for a custom ordinary-workflow step count, for example 普通 50-step. Stop and wait for this answer before assembling and asking the final generation confirmation. For Ref2VA, do not offer this PDD choice: select standard and ask only whether to keep the ordinary 20 steps or use a custom 8–50 step count.
3. A custom step count selects the standard workflow and may be any integer from 8 through 50. PDD-Acc is trained for its locked 8-step path; never run pdd-acc-8step at 20, 50, or another custom count.
4. The PDD preset is the fixed recipe: MiniMax-H3-FL2VA-Acc-8Step.safetensors, Euler sampler, PDD-provided sigmas, NFE 8, LoRA/head strengths 1.0, and MiniMax H3 video/audio sigma shifts 12/3. Do not combine it with Turbo, FirstBlockCache, EasyCache, or another distillation LoRA.