FLUX 3 Image / Klein
This is the canonical local FLUX.3 Klein text-to-image workflow and test
matrix. New reference-based tests belong in optimized/I2I/flux3klein/ and
optimized/MultipleRef2I/flux3klein/; do not create a second T2I directory.
The canonical cross-model comparison row is 1088×1920; the other recorded
resolution rows are explicitly labelled scaling studies.
This directory contains the first controlled workflow and test plan for the
server's locally installed FLUX 3 Image Klein model. The workflow does not use
the remote Flux3ImageNode API.
Workflow
workflow.json— API-format prompt for/promptautomation.prompt.txt— fixed baseline prompt.variables.json— fixed values and allowed test variables.results/— output images and matching metadata records.
The local graph uses flux3-image-klein-int8-convrot.safetensors,
qwen3vl_4b_int8_convrot.safetensors with type=flux3, flux2-vae.safetensors,
Flux guidance, Euler sampling, and the native ComfyUI sampler chain. The
remote API node is not used because it requires a separate Comfy API login.
Test matrix
The first pass changes one variable at a time. The prompt, aspect ratio, grounding state, safety tolerance, and reference count remain fixed unless a row explicitly changes them.
| ID | Purpose | Resolution | References | Grounding | Seed | Status | Runtime (s) | Verdict |
|---|---|---|---|---|---|---|---|---|
| F3I-T2I-01 | Shared-prompt local baseline | 1080×1920 target | 0 | Off | 310001 | Passed | 19.6 | 1088×1920 output generated |
| F3I-T2I-02 | Lower-cost portrait comparison | 0.75K target | 0 | Off | 310002 | Passed | 5.9 | 768×1360 output generated |
| F3I-T2I-03 | Detail/quality portrait comparison | 1.5K target | 0 | Off | 310003 | Passed | 13.9 | 1536×2720 output generated |
| F3I-T2I-04 | Higher-resolution portrait feasibility | 2K target | 0 | Off | 310004 | Passed | 47.2 | 2048×3648 output generated |
| F3I-T2I-05 | Grounded prompt comparison | 1K | 0 | On | 310005 | N/A local | — | Grounding belongs to remote API node |
| F3I-I2I-01 | Single-reference edit | 1K | 1 | Off | 310006 | Passed | 33.6 | 1024×1024 output generated |
Evaluation
Record the exact prompt, seed, resolution, reference filename/hash, runtime,
output dimensions, success/failure, and visual notes in results/<test-id>.json.
Keep the first successful baseline as the comparison target. Do not promote a
larger resolution solely because it is larger; judge detail, composition,
prompt adherence, artifacts, and cost/runtime together.
The API node's seed controls rerun identity but the service may remain nondeterministic. That is recorded as a property of the test, not treated as a failure.
Matrix results
The initial resolution and reference runs were calibration tests with a short
fox prompt. They remain useful for checking graph/model compatibility but are
superseded for model comparison by the shared prompt baseline. The native
reference conditioning path passed with Base001.png. The grounding row is not
run locally because grounding is a capability of the authenticated remote
Flux3ImageNode, which is intentionally excluded from this repository's
workflow.
Initial run
The earlier API-node probe is retained as results/F3I-T2I-01-remote-api.json
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. The analysis and visual
spot-check findings are in 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.
The five strongest visually reviewed outputs are collected in
results/best5/, ordered best first.