From 013ec43bfca1d6791910ce0040ee2ed2978714e1 Mon Sep 17 00:00:00 2001 From: Morpheus Date: Wed, 28 Oct 2026 16:41:17 +0000 Subject: [PATCH] Add deferred Flux 3 Klein test matrix --- optimized/T2I/flux3klein/README.md | 35 +++++++++++++++++++++++++ optimized/T2I/flux3klein/variables.json | 33 +++++++++++++++++++++++ 2 files changed, 68 insertions(+) create mode 100644 optimized/T2I/flux3klein/README.md create mode 100644 optimized/T2I/flux3klein/variables.json diff --git a/optimized/T2I/flux3klein/README.md b/optimized/T2I/flux3klein/README.md new file mode 100644 index 0000000..292969e --- /dev/null +++ b/optimized/T2I/flux3klein/README.md @@ -0,0 +1,35 @@ +# FLUX.3 Klein local test plan + +This plan is prepared for the locally installed `flux3-image-klein-int8-convrot.safetensors` model. Tests are deliberately deferred until **18:30 Europe/London on 28 October 2026**. + +## Available local model candidates + +The ComfyUI server currently reports these text encoders: + +- `qwen3-vl-4b-heretic_int8.safetensors` — primary candidate. The VL 4B family is the closest available match to the Klein workflow's expected Qwen3-VL conditioning architecture. +- `qwen3-8b-heretic_int8.safetensors` — compatibility probe only. It is a text-only Qwen3 family model, so it should not be assumed compatible with the Klein VL conditioning path. +- `qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors` and `qwen3vl_8b_w4a8.safetensors` — excluded from the first matrix because they are H3-oriented variants with different expected conditioning contracts. +- `uncensoredQwen34BLLMForZ_fp8.safetensors` — excluded; not the required Qwen3-VL architecture. + +The model and encoder must be validated by ComfyUI's node contract and actual prompt execution; filename similarity alone is not sufficient. + +## Basic matrix + +All runs use the existing shared prompt from `../flux3image/prompt.txt`, the local Flux 3 Klein diffusion model, the local Flux-compatible VAE, fixed seed `310101`, guidance `3.5`, and four steps unless a failure requires a diagnostic rerun. + +| ID | Encoder | Mode | Resolution | References | Sampler / scheduler | Purpose | +|---|---|---|---:|---:|---|---| +| F3K-T2I-01 | Qwen3-VL 4B Heretic int8 | text-to-image | 1024×1024 | 0 | Euler / simple | Primary architecture and baseline generation | +| F3K-T2I-02 | Qwen3-VL 4B Heretic int8 | text-to-image | 1088×1920 | 0 | Euler / simple | Existing portrait baseline comparison | +| F3K-T2I-03 | Qwen3-VL 4B Heretic int8 | text-to-image | 1024×1024 | 0 | Euler / normal | Scheduler compatibility check | +| F3K-I2I-01 | Qwen3-VL 4B Heretic int8 | image-to-image | 1024×1024 | 1 | Euler / simple | Reference-image conditioning | +| F3K-MI2I-01 | Qwen3-VL 4B Heretic int8 | multi-image-to-image | 1024×1024 | 2 | Euler / simple | Multiple reference conditioning | +| F3K-PROBE-01 | Qwen3 8B Heretic int8 | text-to-image | 512×512 | 0 | Euler / simple | Controlled negative compatibility probe; stop on contract/shape failure | + +## Test rules + +1. Confirm the exact files exposed by `/models/text_encoders`, `/models/diffusion_models`, and `/models/vae` immediately before queueing. +2. Do not substitute H3-specific Qwen3-VL encoders into this matrix without recording the contract mismatch. +3. Queue the primary VL rows first. Do not queue the 8B probe if the ComfyUI node rejects its architecture at graph validation. +4. Record prompt, model filenames, encoder type, seed, resolution, sampler, scheduler, runtime, output dimensions, and the complete error text for failures. +5. Compare successful outputs against the existing FLUX.3 Image baseline and retain only the test records and selected outputs needed for analysis. diff --git a/optimized/T2I/flux3klein/variables.json b/optimized/T2I/flux3klein/variables.json new file mode 100644 index 0000000..afd8736 --- /dev/null +++ b/optimized/T2I/flux3klein/variables.json @@ -0,0 +1,33 @@ +{ + "model": "flux3-image-klein-int8-convrot.safetensors", + "candidate_text_encoders": [ + { + "file": "qwen3-vl-4b-heretic_int8.safetensors", + "type": "flux3", + "role": "primary", + "reason": "Closest available Qwen3-VL 4B architecture to the Klein workflow contract" + }, + { + "file": "qwen3-8b-heretic_int8.safetensors", + "type": "flux3", + "role": "compatibility-probe", + "reason": "Heretic Qwen3 family variant, but text-only and not presumed VL-compatible" + } + ], + "fixed": { + "prompt_file": "../flux3image/prompt.txt", + "seed": 310101, + "guidance": 3.5, + "steps": 4, + "model": "flux3-image-klein-int8-convrot.safetensors", + "vae": "flux2-vae.safetensors" + }, + "tests": [ + {"id":"F3K-T2I-01","encoder":"qwen3-vl-4b-heretic_int8.safetensors","mode":"t2i","width":1024,"height":1024,"references":0,"sampler":"euler","scheduler":"simple"}, + {"id":"F3K-T2I-02","encoder":"qwen3-vl-4b-heretic_int8.safetensors","mode":"t2i","width":1088,"height":1920,"references":0,"sampler":"euler","scheduler":"simple"}, + {"id":"F3K-T2I-03","encoder":"qwen3-vl-4b-heretic_int8.safetensors","mode":"t2i","width":1024,"height":1024,"references":0,"sampler":"euler","scheduler":"normal"}, + {"id":"F3K-I2I-01","encoder":"qwen3-vl-4b-heretic_int8.safetensors","mode":"i2i","width":1024,"height":1024,"references":1,"sampler":"euler","scheduler":"simple"}, + {"id":"F3K-MI2I-01","encoder":"qwen3-vl-4b-heretic_int8.safetensors","mode":"mi2i","width":1024,"height":1024,"references":2,"sampler":"euler","scheduler":"simple"}, + {"id":"F3K-PROBE-01","encoder":"qwen3-8b-heretic_int8.safetensors","mode":"t2i","width":512,"height":512,"references":0,"sampler":"euler","scheduler":"simple"} + ] +}