# Qwen NoctQ v4 → FLUX.3 Klein refiner I2I Two-stage local image-to-image workflow. Qwen NoctQ v4 generates the first image, then the decoded pixels are passed to FLUX.3 Klein as both the image reference and the VAE-encoded latent for a low-denoise refinement pass. The FLUX stage has **no prompt text**; it refines only from the supplied image conditioning and latent. The general ComfyUI sampler path is used so `euler_ancestral` remains available for the FLUX stage. ## Fixed starter settings | Stage | Model | Steps | CFG / guidance | Sampler | Scheduler | Denoise | |---|---|---:|---:|---|---|---:| | Qwen base | `NoctQ_V4_int8_convrot.safetensors` | 8 | CFG 1.0 | `euler_ancestral` | `beta` | 1.0 | | FLUX.3 Klein refiner | `flux3-image-klein-int8-convrot.safetensors` | 8 | Guidance 1.0 | `euler_ancestral` | `simple` | 0.11 | The text encoders are the installed local Heretic variants: `qwen3vl_8b_w4a8.safetensors` for NoctQ and `qwen3-vl-4b-heretic_int8.safetensors` for FLUX.3 Klein. No remote model API is used. The input is the canonical I2I image `../character_reference.png` at 1080×1920 with seed `310101`. The Qwen stage uses the exact shared I2I prompt from `../qwen/workflow.json#7`; the FLUX stage has no text prompt. The Qwen output is decoded with the Qwen VAE, then connected to the FLUX reference conditioner with an empty prompt and re-encoded with `flux2-vae.safetensors` before the 0.11-denoise refiner pass. ## Test matrix | ID | Qwen settings | FLUX.3 Klein settings | Generation time (s) | Status | Review | |---|---|---|---:|---|---| | QNF3K-I2I-01 | 8 steps, CFG 1, euler_ancestral, beta | 8 steps, guidance 1, euler_ancestral, simple, denoise 0.11 | — | Planned | Starter refiner test | Record both stage runtimes, prompt ID, output dimensions, output path, model names, and visual review notes in `results/QNF3K-I2I-01.json` after execution.