200 lines
15 KiB
Markdown
200 lines
15 KiB
Markdown
# MiniMax H3 — optimized reference-to-video
|
||
|
||
This is the MiniMax H3 R2V evaluation workflow. The current production
|
||
baseline is the validated Turbo LoRA graph in `api-workflow.json`; the older
|
||
web-editor graph remains in `workflow.json` for reference and is not the
|
||
current baseline.
|
||
|
||
## Current baseline inputs
|
||
|
||
- Reference: `reference.png`, supplied as Picture 1
|
||
- Prompt: `prompt.txt`
|
||
- Seed: fixed within each comparison batch
|
||
- Resolution: 960×544 landscape
|
||
- Duration: 10 seconds
|
||
- CFG: **1.0**
|
||
- Model, text encoder, VAE, batch size, and denoise: fixed
|
||
- Audio: none
|
||
|
||
The earlier Phase A–O rows include historical 864×480/3-second and
|
||
864×480/10-second comparisons. New tests should use the current 960×544,
|
||
10-second baseline unless a row explicitly specifies another resolution or
|
||
duration.
|
||
|
||
Only the variables in the matrix may change. Before each row is run, confirm
|
||
that the named node or sampler exists on the live server. Unsupported rows
|
||
must be marked `Unsupported` with the exact validation error; do not silently
|
||
substitute another configuration.
|
||
|
||
## Files
|
||
|
||
- `workflow.json` — ComfyUI web workflow derived from the source graph
|
||
- `api-workflow.json` — validated API graph for the current Turbo baseline
|
||
- `reference.png` — fixed character reference
|
||
- `prompt.txt` — fixed bright park-and-robins prompt
|
||
- `variables.json` — fixed comparison variables
|
||
- `results/` — labelled outputs and matching JSON metadata
|
||
|
||
## New sparse-attention latent-upscale variant
|
||
|
||
The supplied Fast H3 sparse-attention and tiled latent-upscale study is
|
||
documented separately in [`sparse-latent-upscale/`](sparse-latent-upscale/).
|
||
It is intentionally not promoted over the current Turbo baseline until the
|
||
controlled upscale, tile and initial-dimension tests are complete.
|
||
|
||
## Current best
|
||
|
||
**MM-R2V-58** remains the historical 3-second baseline: Turbo LoRA enabled, 8
|
||
steps, Res Multistep sampler, simple scheduler, CFG 1, 864×480, and 368.746s
|
||
runtime. **MM-R2V-61** is the quality reference for the original 864×480
|
||
10-second bright park-and-robins prompt. **MM-R2V-68** is now the production
|
||
baseline at 960×544: SLA 0.85 / block 32, Comfy Kitchen attention, Low-VRAM
|
||
Attention (4 chunks), and Chunk FeedForward (2 chunks), with 659.169s runtime.
|
||
|
||
## Test matrix
|
||
|
||
The matrix records historical results, blocked plans, and the current
|
||
960×544 baseline. New tests must use the fixed current-baseline inputs unless
|
||
their row explicitly overrides resolution or duration.
|
||
|
||
| ID | Phase | Efficiency / node combination | Steps | Sampler | Scheduler | Duration (s) | CFG | Tested? | Valid? | Current Best | Runtime (s) |
|
||
|---|---|---|---:|---|---|---:|---:|---|---|---|---:|
|
||
| MM-R2V-01 | A control | Base workflow | 8 | workflow default | workflow default | 3 | 1.0 | Yes | Yes | No | 320.4 |
|
||
| MM-R2V-02 | A memory | MemoryEfficientSageAttentionPatch | 8 | workflow default | workflow default | 3 | 1.0 | Yes | Unsupported | No | 5.0 |
|
||
| MM-R2V-03 | A memory | Model cache/offload path only | 8 | workflow default | workflow default | 3 | 1.0 | Yes | Yes | No | 365.3 |
|
||
| MM-R2V-04 | A attention | Sol-Attn only | 4 | workflow default | workflow default | 3 | 1.0 | Yes | Yes | No | 205.7 |
|
||
| MM-R2V-05 | A attention | Spectrum only | 4 | euler | simple | 3 | 1.0 | Yes | Yes | No | 158.881 |
|
||
| MM-R2V-06 | A LoRA | Turbo LoRA only | 4 | euler | simple | 3 | 1.0 | Yes | Yes | No | 167.078 |
|
||
| MM-R2V-07 | A combined | Memory + Sol-Attn | 4 | euler | simple | 3 | 1.0 | Yes | Yes | No | 202.124 |
|
||
| MM-R2V-08 | A combined | Memory + Spectrum | 4 | euler | simple | 3 | 1.0 | Yes | Yes | No | 148.031 |
|
||
| MM-R2V-09 | A combined | Memory + Turbo LoRA | 4 | euler | simple | 3 | 1.0 | Yes | Yes | No | 179.447 |
|
||
| MM-R2V-10 | A combined | Sol-Attn + Turbo LoRA | 4 | euler | simple | 3 | 1.0 | Yes | Yes | No | 199.032 |
|
||
| MM-R2V-11 | A combined | Spectrum + Turbo LoRA | 4 | euler | simple | 3 | 1.0 | Yes | Yes | No | 146.926 |
|
||
| MM-R2V-12 | A combined | Memory + Sol-Attn + Turbo LoRA | 4 | euler | simple | 3 | 1.0 | Yes | Yes | No | 198.186 |
|
||
| MM-R2V-13 | B steps | Best valid Phase A path | 4 | best valid | best valid | 3 | 1.0 | No | — | No | — |
|
||
| MM-R2V-14 | B steps | Best valid Phase A path | 6 | best valid | best valid | 3 | 1.0 | No | — | No | — |
|
||
| MM-R2V-15 | B steps | Best valid Phase A path | 8 | best valid | best valid | 3 | 1.0 | No | — | No | — |
|
||
| MM-R2V-16 | B steps | Best valid Phase A path | 12 | best valid | best valid | 3 | 1.0 | No | — | No | — |
|
||
| MM-R2V-17 | B sampler | Best valid Phase A path | 6 | euler | simple | 3 | 1.0 | No | — | No | — |
|
||
| MM-R2V-18 | B sampler | Best valid Phase A path | 6 | euler_cfg_pp | simple | 3 | 1.0 | No | — | No | — |
|
||
| MM-R2V-19 | B sampler | Best valid Phase A path | 6 | res_multistep | simple | 3 | 1.0 | No | — | No | — |
|
||
| MM-R2V-20 | B sampler | Best valid Phase A path | 6 | dpmpp_2m | simple | 3 | 1.0 | No | — | No | — |
|
||
| MM-R2V-21 | B scheduler | Best valid Phase A path | 6 | best valid | beta | 3 | 1.0 | No | — | No | — |
|
||
| MM-R2V-22 | B scheduler | Best valid Phase A path | 6 | best valid | sgm_uniform | 3 | 1.0 | No | — | No | — |
|
||
| MM-R2V-23 | B scheduler | Best valid Phase A path | 6 | best valid | karras | 3 | 1.0 | No | — | No | — |
|
||
| MM-R2V-24 | B balance | Best valid Phase A path | 8 | best valid | best valid | 3 | 1.0 | No | — | No | — |
|
||
| MM-R2V-25 | C duration | Best 3-second path | best | best | best | 6 | 1.0 | No | — | No | — |
|
||
| MM-R2V-26 | C duration | Best 3-second path | best | best | best | 10 | 1.0 | No | — | No | — |
|
||
| MM-R2V-27 | C duration | Best 3-second path | best | best | best | 20 | 1.0 | No | — | No | — |
|
||
| MM-R2V-28 | D upscale | Best 3-second path + MiniMax latent upscale | best | best | best | 3 | 1.0 | No | — | No | — |
|
||
| MM-R2V-29 | D upscale | Best 3-second path + RTX Super Resolution | best | best | best | 3 | 1.0 | No | — | No | — |
|
||
| MM-R2V-30 | E extend | Best valid path, concatenate two clips | best | best | best | 6 | 1.0 | No | — | No | — |
|
||
| MM-R2V-31 | E extend | Best valid path, concatenate three clips | best | best | best | 9 | 1.0 | No | — | No | — |
|
||
| MM-R2V-32 | F quality threshold | 20-step graph; Sol-Attn branch present, Spectrum disabled | 20 | er_sde | simple | 3 | 1.0 | Yes | Yes | No | 814.3 |
|
||
| MM-R2V-33 | F quality threshold | 20-step cache graph; Sol-Attn branch present, Spectrum disabled | 20 | er_sde | simple | 3 | 1.0 | Yes | Yes | No | 774.2 |
|
||
| MM-R2V-34 | F quality threshold | Turbo LoRA only, minimum Turbo steps | 4 | euler | simple | 3 | 1.0 | Yes | Yes | No | 220.5 |
|
||
| MM-R2V-35 | F quality threshold | Spectrum + Turbo LoRA, minimum Turbo steps | 4 | euler | simple | 3 | 1.0 | Yes | Yes | No | 145.6 |
|
||
| MM-R2V-36 | G Spectrum sweep | Spectrum only, blend weight 0.25 | 20 | euler | simple | 3 | 1.0 | Yes | Yes | No | 461.9 |
|
||
| MM-R2V-37 | G Spectrum sweep | Spectrum only, blend weight 0.75 | 20 | euler | simple | 3 | 1.0 | Yes | Yes | No | 452.3 |
|
||
| MM-R2V-38 | G Spectrum sweep | Spectrum only, warmup/tail steps 2/2 | 20 | euler | simple | 3 | 1.0 | Yes | Yes | No | 488.0 |
|
||
| MM-R2V-39 | G Spectrum sweep | Spectrum only, max history 16 | 20 | euler | simple | 3 | 1.0 | Yes | Yes | No | 433.8 |
|
||
| MM-R2V-40 | G Spectrum + Turbo | Spectrum + Turbo, blend weight 0.25 | 4 | euler | simple | 3 | 1.0 | Yes | Yes | No | 184.9 |
|
||
| MM-R2V-41 | G Spectrum + Turbo | Spectrum + Turbo, blend weight 0.75 | 4 | euler | simple | 3 | 1.0 | Yes | Yes | No | 181.4 |
|
||
| MM-R2V-42 | G Spectrum + Turbo | Spectrum + Turbo, replay disabled | 4 | euler | simple | 3 | 1.0 | Yes | Yes | No | 145.9 |
|
||
| MM-R2V-43 | G Spectrum + Turbo | Spectrum + Turbo, max history 16 | 4 | euler | simple | 3 | 1.0 | Yes | Yes | No | 149.1 |
|
||
| MM-R2V-44 | H quality sweep | Cache/offload + Sol-Attn | 20 | er_sde | simple | 3 | 1.0 | Blocked | — | No | — |
|
||
| MM-R2V-45 | H quality sweep | Cache/offload + Sol-Attn | 24 | er_sde | simple | 3 | 1.0 | Blocked | — | No | — |
|
||
| MM-R2V-46 | H quality sweep | Cache/offload + Sol-Attn | 28 | er_sde | simple | 3 | 1.0 | Blocked | — | No | — |
|
||
| MM-R2V-47 | H quality sweep | Cache/offload + Sol-Attn | 32 | er_sde | simple | 3 | 1.0 | Blocked | — | No | — |
|
||
| MM-R2V-48 | H quality sweep | Base/Sol-Attn path | 20 | er_sde | simple | 3 | 1.0 | Blocked | — | No | — |
|
||
| MM-R2V-49 | H quality sweep | Base/Sol-Attn path | 24 | er_sde | simple | 3 | 1.0 | Blocked | — | No | — |
|
||
| MM-R2V-50 | H quality sweep | Base/Sol-Attn path | 28 | er_sde | simple | 3 | 1.0 | Blocked | — | No | — |
|
||
| MM-R2V-51 | H Turbo sweep | Turbo LoRA only | 8 | euler | simple | 3 | 1.0 | Yes | Yes | No | 381.240 |
|
||
| MM-R2V-52 | H Turbo sweep | Turbo LoRA only | 12 | euler | simple | 3 | 1.0 | Yes | Yes | No | 471.862 |
|
||
| MM-R2V-53 | H Turbo sweep | Spectrum + Turbo LoRA | 8 | euler | simple | 3 | 1.0 | Blocked | — | No | — |
|
||
| MM-R2V-54 | H Turbo sweep | Spectrum + Turbo LoRA | 12 | euler | simple | 3 | 1.0 | Blocked | — | No | — |
|
||
| MM-R2V-55 | I Turbo sampler/scheduler | Turbo LoRA only | 8 | euler | beta | 3 | 1.0 | Yes | Yes | No | 372.061 |
|
||
| MM-R2V-56 | I Turbo sampler/scheduler | Turbo LoRA only | 8 | euler_cfg_pp | simple | 3 | 1.0 | Yes | Yes | No | 432.718 |
|
||
| MM-R2V-57 | I Turbo sampler/scheduler | Turbo LoRA only | 8 | res_multistep | beta | 3 | 1.0 | Yes | Yes | No | 369.694 |
|
||
| MM-R2V-58 | I Turbo sampler/scheduler | Turbo LoRA only | 8 | res_multistep | simple | 3 | 1.0 | Yes | Yes | Yes | 368.746 |
|
||
| MM-R2V-59 | J Turbo step comparison | Turbo LoRA only | 12 | res_multistep | beta | 3 | 1.0 | Yes | Yes | No | 551.740 |
|
||
| MM-R2V-60 | J Turbo step comparison | Turbo LoRA only | 12 | res_multistep | simple | 3 | 1.0 | Yes | Yes | No | 511.753 |
|
||
| MM-R2V-61 | K duration/prompt | Current Turbo baseline with bright park prompt | 8 | res_multistep | simple | 10 | 1.0 | Yes | Yes | No | 790.632 |
|
||
| MM-R2V-62 | L SLA attention | MM-R2V-61 + H3 SLA Attention, sparsity 0.85, block 32 | 8 | res_multistep | simple | 10 | 1.0 | Yes | Yes | No | 525.745 |
|
||
| MM-R2V-63 | L SLA attention | MM-R2V-61 + H3 SLA Attention, sparsity 0.90, block 32 | 8 | res_multistep | simple | 10 | 1.0 | Yes | Yes | No | 480.653 |
|
||
| MM-R2V-64 | M combined attention | SLA 0.85 / block 32 + Sol-Attention | 8 | res_multistep | simple | 10 | 1.0 | Yes | Yes | No | 597.210 |
|
||
| MM-R2V-65 | M combined attention | SLA 0.85 / block 32 + MiniMax Low-VRAM Attention, 4 head chunks | 8 | res_multistep | simple | 10 | 1.0 | Yes | Yes | No | 552.801 |
|
||
| MM-R2V-66 | N cache | SLA 0.85 / block 32 + EasyCache defaults (reuse 0.2, 15–95%) | 8 | res_multistep | simple | 10 | 1.0 | Yes | Yes | No | 532.431 |
|
||
| MM-R2V-67 | O cleanup control | Clean single-reference graph + SLA 0.85 / block 32 | 8 | res_multistep | simple | 10 | 1.0 | Yes | Yes | No | 538.369 |
|
||
| MM-R2V-68 | P high resolution | 960×544 + SLA 0.85 + Comfy Kitchen backend + Low-VRAM 4 + Chunk FF 2 | 8 | res_multistep | simple | 10 | 1.0 | Yes | Yes | Yes | 659.169 |
|
||
| MM-R2V-69 | Q H3 LongVideos | Two fixed 10-second shots: robins, then squirrel joins; full baseline attention/memory chain | 8 | res_multistep | simple | 20 | 1.0 | Yes | Blocked | No | — |
|
||
| MM-R2V-70 | R model comparison | Current MM-R2V-68 graph with Singularity checkpoint | 8 | res_multistep | simple | 10 | 1.0 | Yes | Yes | No | 879.775 |
|
||
| MM-R2V-71 | R model comparison | Current MM-R2V-68 graph with standard ref2va checkpoint | 8 | res_multistep | simple | 10 | 1.0 | Yes | Yes | No | 604.242 |
|
||
|
||
## Result requirements
|
||
|
||
Each run must save `results/<test-id>.mp4` and matching
|
||
`results/<test-id>.json`. Metadata must include the exact reference checksum,
|
||
prompt hash, seed, resolution, duration, CFG, steps, sampler, scheduler,
|
||
active nodes, runtime, VRAM, output checksum, validity, and visual/motion
|
||
verdict. A result becomes Current Best only when it is valid and objectively
|
||
better than the existing best; keep exactly one Current Best.
|
||
|
||
Phase A collection status: rows 01 and 03–12 have valid MP4 outputs recorded; MM-R2V-05 through MM-R2V-12 were collected from the 2026-09-26 live ComfyUI history. Row 02 remains unsupported because the live SageAttention patch failed validation/runtime. Subjective visual review was not performed, so Current Best was not changed.
|
||
|
||
Phase F rule: non-Turbo configurations must use at least 20 steps; Turbo
|
||
LoRA configurations may use 4–12 steps, with 12 as the current upper limit
|
||
until there is comparative evidence for higher values. This rule is applied
|
||
to the queued quality-threshold and Spectrum tests above.
|
||
|
||
Phase F/G collection status: MM-R2V-32 through MM-R2V-43 all completed with
|
||
valid MP4 outputs. Technical evaluation and runtime capture are complete;
|
||
subjective frame-by-frame quality review is still pending. The 20-step
|
||
non-Turbo graph inherited disabled Spectrum and present Sol-Attn branches,
|
||
which is recorded in the matching metadata.
|
||
|
||
Rows MM-R2V-44 through MM-R2V-50 and MM-R2V-53 through MM-R2V-54 are blocked
|
||
until the corresponding cache/offload, Sol-Attn, and Spectrum branches are
|
||
present in the live API graph. They must not be substituted with the plain
|
||
Turbo/base graph. MM-R2V-51 through MM-R2V-58 completed with the validated
|
||
Turbo graph. User review promoted MM-R2V-58 for its environment quality and
|
||
lower runtime; MM-R2V-59/60 did not justify the additional 12-step runtime.
|
||
MM-R2V-61 was user-reviewed as clean with an acceptable reference match.
|
||
MM-R2V-62 was judged acceptable but lower quality than MM-R2V-61, and was
|
||
retained as the earlier 864×480 speed comparison. MM-R2V-63 was judged
|
||
acceptable but not as good as MM-R2V-62, so it was not promoted.
|
||
MM-R2V-68 was user-reviewed as having very good quality and is promoted as
|
||
the current baseline at 960×544. The current API graph and variables now
|
||
match this result.
|
||
MM-R2V-64 was user-reviewed as having no visible quality difference from the
|
||
SLA-only baseline, with the added Sol-Attention cost, so it was not promoted.
|
||
MM-R2V-65 was user-reviewed as having no clear quality difference from the
|
||
SLA-only baseline, so it was not promoted.
|
||
MM-R2V-66 was user-reviewed as quality-consistent with the SLA-only baseline.
|
||
MM-R2V-67 was user-reviewed as quality-consistent with the previous SLA
|
||
baseline, confirming that removing the duplicate reference and unused nodes
|
||
did not change the result quality.
|
||
|
||
MM-R2V-70 and MM-R2V-71 completed successfully as a checkpoint comparison
|
||
using the MM-R2V-68 graph. MM-R2V-70 used the Singularity checkpoint and ran
|
||
in 879.775 seconds; MM-R2V-71 used the standard ref2va checkpoint and ran in
|
||
604.242 seconds. Both outputs are valid and awaiting subjective visual review;
|
||
neither is promoted until the user compares quality and reference fidelity.
|
||
|
||
MM-R2V-69 is blocked by the live server's missing H3-compatible text encoder.
|
||
The available `qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors` emits 5120
|
||
features, while H3 LongVideos requires the 2560-feature
|
||
`qwen3vl_32b_h3_ultra_uncensored_heretic_int8_convrot.safetensors` referenced
|
||
by the imported workflow. Do not substitute another encoder; install or
|
||
restore the H3-compatible model before rerunning.
|
||
|
||
## Execution order
|
||
|
||
1. Use MM-R2V-68 as the current 960×544/10-second baseline.
|
||
2. Compare new attention, cache, and sampler changes against MM-R2V-68.
|
||
3. Keep prompt, reference, seed, CFG, duration, and resolution fixed unless a
|
||
test row explicitly changes them.
|
||
4. After the baseline is stable, test the installed H3 LongVideos workflow
|
||
for paragraph-driven shot chaining, then latent upscale and RTX Super
|
||
Resolution.
|