Files
comfyui-workflows/optimized/I2I/flux3klein

FLUX.3 Klein image-to-image tests

This is the FLUX.3 Klein image-to-image section. Keep one-reference tests here rather than in the text-to-image directory.

The canonical comparison fixture is ../character_reference.png, and the canonical prompt is the shared I2I prompt in ../flux/workflow.json (node 74). These inputs must remain fixed when comparing FLUX.3 Klein with other models; only the tested generation settings may vary.

The cross-model output target is fixed at 1080×1920 (the source fixture is 736×1308). Resolution changes are out of scope for this comparison matrix. CFG is fixed at 1.0 and steps at 16 for the sampler matrix. The first follow-up variance is a CFG sweep using the completed F3K-I2I-01 settings as the control and changing only CFG.

Sampler matrix

F3K-I2I-01 completed locally. The matrix keeps the source image, prompt, seed, dimensions, CFG, steps, model, encoder, and scheduler fixed while varying only the sampler. The local I2I graph uses Flux2Scheduler, which does not expose the selectable scheduler family used by the T2I BasicScheduler matrix.

ID Encoder Resolution References Seed CFG Steps Sampler Scheduler Generation time (s) Status Purpose
F3K-I2I-01 qwen3-vl-4b-heretic_int8.safetensors 1080×1920 requested; 1072×1920 actual 1 310101 1.0 16 euler Flux2Scheduler 46.339 Completed Baseline; 16-pixel latent quantization reduced width
F3K-I2I-02 qwen3-vl-4b-heretic_int8.safetensors 1080×1920 1 310101 1.0 16 euler_cfg_pp Flux2Scheduler — Planned CFG++ comparison
F3K-I2I-03 qwen3-vl-4b-heretic_int8.safetensors 1080×1920 1 310101 1.0 16 euler_ancestral Flux2Scheduler — Planned Ancestral comparison
F3K-I2I-04 qwen3-vl-4b-heretic_int8.safetensors 1080×1920 1 310101 1.0 16 euler_ancestral_cfg_pp Flux2Scheduler — Planned Ancestral CFG++ comparison
F3K-I2I-05 qwen3-vl-4b-heretic_int8.safetensors 1080×1920 1 310101 1.0 16 dpmpp_2m Flux2Scheduler — Planned DPM++ comparison
F3K-I2I-06 qwen3-vl-4b-heretic_int8.safetensors 1080×1920 1 310101 1.0 16 dpmpp_2m_cfg_pp Flux2Scheduler — Planned DPM++ CFG++ comparison
F3K-I2I-07 qwen3-vl-4b-heretic_int8.safetensors 1080×1920 1 310101 1.0 16 dpmpp_sde Flux2Scheduler — Planned DPM++ SDE comparison
F3K-I2I-08 qwen3-vl-4b-heretic_int8.safetensors 1080×1920 1 310101 1.0 16 heun Flux2Scheduler — Planned Heun comparison
F3K-I2I-09 qwen3-vl-4b-heretic_int8.safetensors 1080×1920 1 310101 1.0 16 res_multistep Flux2Scheduler — Planned RES multistep comparison
F3K-I2I-10 qwen3-vl-4b-heretic_int8.safetensors 1080×1920 1 310101 1.0 16 res_multistep_cfg_pp Flux2Scheduler — Planned RES multistep CFG++ comparison
F3K-I2I-11 qwen3-vl-4b-heretic_int8.safetensors 1080×1920 1 310101 1.0 16 ddim Flux2Scheduler — Planned DDIM comparison
F3K-I2I-12 qwen3-vl-4b-heretic_int8.safetensors 1080×1920 1 310101 1.0 16 uni_pc Flux2Scheduler — Planned UniPC comparison

CFG variance matrix

This sweep reuses the F3K-I2I-01 source image, prompt, seed, resolution, model, encoder, VAE, 16 steps, Euler sampler, and Flux2Scheduler. Only CFG changes. The baseline at CFG 1.0 is already recorded as F3K-I2I-01; run the following values in ascending order through CFG 5.0.

ID Based on CFG Steps Sampler Scheduler Generation time (s) Status Purpose
F3K-I2I-CFG-01 F3K-I2I-01 1.5 16 euler Flux2Scheduler — Planned CFG +0.5
F3K-I2I-CFG-02 F3K-I2I-01 2.0 16 euler Flux2Scheduler — Planned CFG +1.0
F3K-I2I-CFG-03 F3K-I2I-01 2.5 16 euler Flux2Scheduler — Planned CFG +1.5
F3K-I2I-CFG-04 F3K-I2I-01 3.0 16 euler Flux2Scheduler — Planned CFG +2.0
F3K-I2I-CFG-05 F3K-I2I-01 3.5 16 euler Flux2Scheduler — Planned CFG +2.5
F3K-I2I-CFG-06 F3K-I2I-01 4.0 16 euler Flux2Scheduler — Planned CFG +3.0
F3K-I2I-CFG-07 F3K-I2I-01 4.5 16 euler Flux2Scheduler — Planned CFG +3.5
F3K-I2I-CFG-08 F3K-I2I-01 5.0 16 euler Flux2Scheduler — Planned CFG +4.0

Use ../character_reference.png, the canonical I2I prompt, and the installed flux3-image-klein-int8-convrot.safetensors. Record the reference path/hash, runtime, output dimensions, output path, visual result, and exact error for every row. A rejected sampler is a recorded result, not a reason to silently substitute another sampler.