# MiniMax Notes

## Current default assumptions

- `MiniMax H3` is promising, but brute-force paths are too heavy to be the first bet on a `3070`.
- Speed-first and low-VRAM community workflows matter more than official purity for the first real test.

## Strong current source workflows

- `models/minimax/minimax-h3-official-t2v/`
- `models/minimax/minimax-h3-official-i2v/`
- `models/minimax/minimax-h3-turbo-lora-community/`
- `models/minimax/minimax-h3-int8-r2v-community/`
- `models/minimax/minimax-h3-extender-ref2va/`
- `models/minimax/minimax-h3-int8-r2v-javano2608-23/`
- `models/minimax/minimax-h3-turbo-gguf-i2va/`
- `models/minimax/minimax-h3-seamless-chain-v2/`

## Strong current heuristics

- Prefer Turbo LoRA for the first `text-to-video` attempt.
- Prefer INT8-based community workflows for `reference-to-video` exploration.
- Keep attention and memory-efficiency helpers if they are already integrated and well documented.
- Treat the same-day `javawock7618` `INT8 R2V` import as the strongest current donor for local continuity and identity retention, but not as proof that the graph is light enough for a `3070`.
- Keep the GGUF `I2VA` path as an audio-aware fallback, not the default lane.

## Current weak areas

- No proof yet that MiniMax is truly practical on Chris's box.
- Custom-node stack is more fragile than the current `flux2klein` path.
- The full `seamless-chain` graph is interesting, but it is not a clean baseline for the first maintained master.

## Best current optimization direction

- Keep `optimized/text-to-video/minimax-h3-3070-turbo-master/` as the first MiniMax default.
- Keep the INT8 reference-video workflow as the next donor for future low-VRAM optimization.
- Keep only two maintained `optimized/image-to-video/` MiniMax lanes:
  - `optimized/image-to-video/minimax-h3-best-current-r2v-master-480p-3x10s/` as the reference-first generation master
  - `optimized/image-to-video/minimax-h3-contex-loop-3070-safe-60s-awq-int8/` as the imported-video extension master
- Use the new same-day `javawock7618` `INT8 R2V` import as the main donor for future `best-current-r2v` revisions if the current master needs another pass.

## Ref2VA prompt guidance

- For `MiniMax H3 Ref2VA`, treat `480p` and `10s` as the safe planning baseline on `8GB VRAM` unless direct tests prove otherwise.
- Chris's later manual testing suggests `20s @ 480p` is a real context-loop candidate ceiling, so maintained loop masters should compare `3 x 20s` against `6 x 10s` rather than assuming only `10s` clips are viable.
- In the maintained `best-current-r2v-master`, keep `Picture 1` as the master scene / environment reference and `Picture 2` as the primary full-body identity reference.
- Use one master-scene reference to control environment/composition, one primary full-body reference per subject to control overall appearance, and headshots only to reinforce facial identity.
- Explicitly state when multiple images represent the same subject, and explicitly tell the model to ignore character-reference backgrounds so they do not fight the master environment.
- Use the prompt mainly for action, performance, camera movement, dialogue, and audio rather than re-describing scene details already visible in the references.
- Full guidance now lives in `optimized/notes/minimax-h3-ref2va-prompt-guidelines.md`.

## Loop enhancement direction

- For the target maintained `Contex Loop` quality workflow, treat the second-pass enhancement lane as mandatory, not optional.
- That second pass should reuse the conditioning from the original `MiniMaxH3ReferenceToVideo` node that gathers the reference images, and feed that original reference conditioning back into the guider for the enhancement pass instead of relying only on first-pass latent continuity.
- Current preferred sampler/scheduler direction for that enhancement pass is `Euler` plus `beta`.

## What needs Monday proof

- whether the Turbo LoRA workflow is just technically possible or actually usable
- whether the source-recommended `8`-step path holds up in quality
- whether MiniMax feels like a real challenger to `LTX` locally
- whether the same-day `INT8 R2V` donor stays practical on the `3070` once the helper stack is trimmed to the minimum