MiniMax H3 sparse-attention latent-upscale study
This workflow variant is based on the supplied CSGlide H3 workflow. It uses:
fastvideo_fasth3_8step_v2_pruned_int8_convrot.safetensors- Block Sparse Attention with VSA selection
- Comfy Kitchen attention, MiniMax low-VRAM attention and chunked feed-forward
- 8-step Euler + beta initial generation
- MiniMax latent upscaling with tiled spatial splitting and latent refinement
- The supplied VAE, audio VAE, references, seed and one approximately 10-second clip
The workflow is preserved in workflow-source.json; workflow.json is the tracked test copy with the enhanced common prompt and corrected reference filenames.
For the visual ComfyUI editor, load workflow-ui.json. It is the canvas-format version of the same validated baseline, with node positions and links. The API-format workflow.json remains the file used by the automated submission scripts.
Fixed test input
- References:
references/Scene022.pngas@image1,references/Char034_TurnaroundFull_001.pngas@image2 - Clip: 243 frames at 24 fps, approximately 10.125 seconds
- Seed:
317182692766673for the initial sampler; keep it unchanged in every test - Initial sampler: Euler, beta scheduler, 8 steps, denoise 1.0
- Latent refinement: SA solver, simple scheduler, 3 steps, denoise 0.2, CFG 1
- Prompt:
prompt.txt; do not change it between tests
Test matrix
Every test changes only the listed variables. Record runtime, peak VRAM, success/failure, output dimensions, visible seams, detail retention, anatomy, motion stability and audio output in the matching result JSON.
Phase A — Upscale ceiling
| ID | Initial latent | Target latent | Tile | Purpose |
|---|---|---|---|---|
| SLU-01 | 672×384 | 1344×768 | supplied baseline tile | Reproduce the supplied good result |
| SLU-02 | 672×384 | 1920×1080 | supplied baseline tile | Test the desired 1080p endpoint |
| SLU-03 | 672×384 | 1680×960 | supplied baseline tile | Intermediate fallback if 1080p fails or is too slow |
| SLU-04 | 672×384 | 1536×864 | supplied baseline tile | Practical quality/runtime fallback |
Phase B — Tile size
Use the best successful target from Phase A and compare tile sizes while keeping overlap and all other settings fixed:
- SLU-05: 512×512 tiles
- SLU-06: 640×640 tiles
- SLU-07: 768×768 tiles
- SLU-08: 864×480-style wide tiles from the supplied workflow
- SLU-09: full-frame/no split if VRAM permits
For each tile, preserve 128-pixel spatial overlap where possible, then check seam visibility, runtime and peak VRAM.
Phase C — Initial generation dimensions
Use the best target and tile combination:
- SLU-10: 672×384 supplied baseline
- SLU-11: 768×432
- SLU-12: 832×480 supplied alternate resolution node
- SLU-13: 960×544
- SLU-14: 1024×576 if the 8GB card remains stable
The initial aspect ratio must remain 16:9 and the same final target must be used for fair comparison.
Phase D — Combined candidates
- SLU-15: best upscale ceiling + best tile + best initial dimensions
- SLU-16: quality-first combination, accepting a longer runtime if the visual improvement is clear
- SLU-17: practical production combination targeting a predictable sub-hour render
- SLU-18: rectangular 1024×640 tile shape with four effective tile passes
- SLU-19: 96px spatial overlap on the accepted 1024×640 rectangular tile layout
First two runs
Run SLU-01 and SLU-02 first. SLU-01 validates the supplied configuration. SLU-02 directly tests whether 1080p is feasible on the RTX 3070. If SLU-02 exhausts VRAM, stalls or becomes impractically slow, stop that run and continue with SLU-03/04 rather than changing the prompt or seed.
Evidence and verdict
The test is VALIDATED only when a configuration completes with a usable video and recorded measurements. A completed render with unacceptable seams, detail loss or motion instability is PARTIAL, not a success. Keep all result JSON and MP4 files under results/.
Recorded valid runs
These are the valid runs recorded so far. Runtime is wall-clock generation time; the normalized figure is included because the output sizes differ.
| ID | Initial → output | Tile | Runtime | Runtime per output megapixel | Output | Review |
|---|---|---|---|---|---|---|
| SLU-01 | 672×384 → 1344×768 | 864×480 | 20m 09s | 19.51 min/Mpx | 1344×768 | Pending |
| SLU-02 | 672×384 → 1920×1088 | 864×480 | 32m 06s | 15.37 min/Mpx | 1920×1088 | Baseline |
| SLU-05 | 672×384 → 1920×1088 | 512×512 | 44m 48s | 21.45 min/Mpx | 1920×1088 | Valid; slower |
| SLU-06 | 672×384 → 1920×1088 | 640×640 | 26m 52s | 12.88 min/Mpx | 1920×1088 | Reviewed; good |
| SLU-07 | 672×384 → 1920×1088 | 768×768 | 26m 01s | 12.45 min/Mpx | 1920×1088 | Current best |
| SLU-08 | 672×384 → 1920×1088 | 864×480 | 31m 13s | 14.93 min/Mpx | 1920×1088 | Valid; visual comparison pending |
| SLU-11 | 768×432 → 1920×1088 | 768×768 | 31m 07s | 14.91 min/Mpx | 1920×1088 | Valid; review pending |
| SLU-18 | 672×384 → 1920×1088 | 1024×640 | 25m 14s | 12.08 min/Mpx | 1920×1088 | Reviewed; new current best |
| SLU-19 | 672×384 → 1920×1088 | 1024×640, 96px overlap | 18m 11s | 8.71 min/Mpx | 1920×1088 | Reviewed; accepted current best |
Current best
SLU-19 is now the current best baseline: 672×384 initial generation, 1024×640 tiles in a 2×2 layout, 96-pixel overlap and a 1920×1088 output. It was reviewed and accepted with good quality and no visible seams. At 18m 11s it is below the 20-minute target and has the best normalized throughput so far at 8.71 min/Mpx. The attention keep_percent 5 test was slower at 31m 05s, so the baseline value of 10 remains in use. SLU-18 remains the same tile shape with 128px overlap at 25m 14s.
SLU-09 full-frame/no-split failed with GPU out-of-memory on the 8 GB card and is not a viable baseline. The tracked workflow.json now uses the SLU-18 settings. Static prompt, seed, references, model, samplers, schedulers, steps, denoise and final target remain unchanged.
The earlier tile submissions that produced 1024×1024 output are excluded from this table because their target-resolution override was not applied. The current baseline uses 96px overlap; further reductions should only be attempted if a visible-seam comparison is worthwhile.