Record Flux 3 sampler scheduler matrix

This commit is contained in:
Morpheus
2026-10-22 21:10:40 +00:00
parent d5c46c5496
commit 454da2d645
3 changed files with 2420 additions and 0 deletions
@@ -0,0 +1,72 @@
# FLUX 3 Image sampler/scheduler matrix
## Test conditions
- **Runs:** 432 (48 samplers × 9 schedulers)
- **Model:** `flux3-image-klein-int8-convrot.safetensors`
- **Text encoder:** `qwen3vl_4b_int8_convrot.safetensors` (`type=flux3`)
- **VAE:** `flux2-vae.safetensors`
- **Prompt:** [`prompt.txt`](prompt.txt), SHA-256 `5802909ebe894260b70a3977430e08aaa95732e0559749d8cb42e2f5ecf3ab6f`
- **Output:** 1088 × 1920 (the 1080 × 1920 target snapped to the model-compatible width)
- **Steps:** 4
- **Seed:** 310001
- **Result:** 432/432 successful
- **Total execution time:** 3726.104 seconds
- **Mean execution time:** 8.625 seconds
The complete per-cell record is in [`results/F3I-MATRIX.json`](results/F3I-MATRIX.json). Each cell contains the status, measured runtime, and the `F3I-M-SS-DD` output prefix used by the server, where `SS` is the sampler index and `DD` is the scheduler index.
## Runtime findings
Fastest sampler means across all nine schedulers:
- `TBG Flux2 Sampler` — 6.257 s
- `pid_sde` — 6.359 s
- `euler` — 6.407 s
- `dpm_fast` — 6.408 s
- `res_multistep` — 6.412 s
- `cfgpp_ud10_ab` — 6.415 s
- `ipndm` — 6.417 s
- `ipndm_v` — 6.424 s
- `ddim` — 6.428 s
- `dpmpp_2m` — 6.429 s
Scheduler means across all 48 samplers:
- `sgm_uniform` — 8.022 s
- `simple` — 8.063 s
- `linear_quadratic` — 8.265 s
- `beta` — 8.374 s
- `kl_optimal` — 8.816 s
- `karras` — 8.829 s
- `exponential` — 8.919 s
- `normal` — 8.921 s
- `ddim_uniform` — 9.417 s
`dpm_adaptive` is a clear runtime outlier. Its slowest cells were:
- `dpm_adaptive` + `ddim_uniform` — 87.590 s
- `dpm_adaptive` + `normal` — 61.760 s
- `dpm_adaptive` + `exponential` — 61.244 s
- `dpm_adaptive` + `kl_optimal` — 57.667 s
- `dpm_adaptive` + `karras` — 57.582 s
The fastest individual cell was `euler` + `exponential` at 6.199 s. Runtime alone is not a quality score.
## Visual spot-check
Representative generated images were inspected after the automated run:
- `F3I-M-01-01` (`euler` + `simple`) was a strong baseline: coherent subject, composition, anatomy, and useful prompt adherence.
- `F3I-M-01-04` (`euler` + `exponential`) was visibly degraded in the inspected sample, with garbled scoreboard text and smeared detail.
- `F3I-M-16-05` (`dpm_adaptive` + `ddim_uniform`) showed severe anatomy duplication/extra limbs, garbled text, and texture artifacts.
- `F3I-M-48-09` (`uni_pc_bh2` + `kl_optimal`) was overprocessed in the inspected sample, with distorted texture/anatomy and garbled text.
These are representative spot-checks, not a claim that four images fully rank all 432 outputs. The matrix proves that every combination executes successfully on the installed local model; a full perceptual ranking would require reviewing all 432 images or adding an image-quality scoring pass.
## Recommendation
Keep **Euler + Simple** as the default comparison baseline. It is fast, already aligned with the native workflow path, and produced the strongest inspected sample. Treat `sgm_uniform` as a similarly fast scheduler candidate for follow-up visual testing. Avoid `dpm_adaptive` for routine generation because its runtime is unpredictable and its inspected sample was poor. Keep the remaining combinations available as experimental options until they receive visual review.
No generated server files were deleted during collection. The matrix outputs remain under the server prefix `fluxExperiments/flux3image/matrix/` so the visual findings can be revisited.
+13
View File
@@ -60,3 +60,16 @@ as an environment diagnostic: it returned `Unauthorized` before generation.
The local workflow was then corrected to use the Flux 3-compatible
Qwen3-VL 4B encoder and 128-channel `flux2-vae.safetensors`; that run passed
and is recorded in `results/F3I-T2I-01.json`.
## Full sampler/scheduler matrix
The local model was tested with every available sampler and scheduler: 48 × 9
combinations, for 432 total runs. All 432 completed successfully at 1088 ×
1920, four steps, seed 310001, using the shared prompt. The complete cell-level
results, runtimes, and server output prefixes are recorded in
[`results/F3I-MATRIX.json`](results/F3I-MATRIX.json). The analysis and visual
spot-check findings are in [`MATRIX-ANALYSIS.md`](MATRIX-ANALYSIS.md).
The recommended baseline remains **Euler + Simple**. `dpm_adaptive` is not
recommended for routine use because it was an extreme runtime outlier; runtime
success does not by itself establish visual quality.
File diff suppressed because it is too large Load Diff