51 lines
2.2 KiB
Markdown
51 lines
2.2 KiB
Markdown
# MiniMax H3 — optimized reference-to-video
|
||
|
||
This is the first cleaned MiniMax H3 R2V evaluation workflow. It is derived
|
||
from the compact official MiniMax H3 int8 R2V graph and is kept separate from
|
||
the imported source workflows under `models/minimax/`.
|
||
|
||
## Fixed variables
|
||
|
||
- Reference media: `reference.png`, supplied to the workflow as Picture 1
|
||
- Prompt: `prompt.txt`, fixed for this comparison batch
|
||
- Seed: keep fixed for each comparison batch
|
||
- Resolution: 864×480 (480p landscape)
|
||
- Duration: 3 seconds for the initial efficiency matrix
|
||
- CFG: **1.0**
|
||
- Model, text encoder, VAE, batch size, and denoise: fixed
|
||
|
||
## Base workflow
|
||
|
||
- `workflow.json` — derived ComfyUI web workflow
|
||
- `reference.png` — fixed character reference for the initial batch
|
||
- `prompt.txt` — fixed haunted-room investigation prompt
|
||
- Source control graph: `models/minimax/minimax-h3-int8-r2v/workflow.json`
|
||
- The imported source is not modified.
|
||
|
||
## Efficiency test matrix
|
||
|
||
The first two tests compare a normal control against the first compatible
|
||
memory-saving patch. All rows use the fixed variables above. The initial test
|
||
does not vary steps, sampler, or scheduler so that the efficiency difference
|
||
is isolated.
|
||
|
||
| ID | Memory path | Steps | Sampler | Scheduler | Duration | CFG | Runtime (s) | VRAM | Tested? | Valid? | Current Best |
|
||
|---|---|---:|---|---|---:|---:|---:|---:|---|---|---|
|
||
| MM-R2V-01 | Normal/control | 8 | workflow default | workflow default | 3 | 1.0 | — | — | No | — | No |
|
||
| MM-R2V-02 | MiniMaxH3MemoryEfficientSageAttentionPatch | 8 | workflow default | workflow default | 3 | 1.0 | — | — | No | — | No |
|
||
|
||
## Planned expansion
|
||
|
||
After these two tests:
|
||
|
||
1. Test other compatible memory nodes individually.
|
||
2. Test safe combinations of compatible nodes.
|
||
3. Sweep steps, sampler, and scheduler with the best efficiency path.
|
||
4. Run only the strongest 3-second configurations at 20 seconds.
|
||
5. Test reference-driven continuation/concatenation.
|
||
6. Compare latent upscale and RTX Video Super Resolution toward 1080p.
|
||
|
||
Every result must be saved as `results/<test-id>.*` with matching JSON
|
||
metadata containing the exact reference, prompt, seed, resolution, duration,
|
||
CFG, node settings, runtime, VRAM, validity, and visual verdict.
|