FLUX.3 [klein]

Highlights

• Text-to-image and instruction editing in one model; editing can take multiple reference images (<ref_image_0>, <ref_image_1>…)
• Two checkpoints: full version at 50 steps, CFG 3.5; distilled version at 16 steps with CFG fixed at 1 (adjustable from 4 to 16 steps)
• Bundled Qwen3.5-9B prompt enhancer: a short prompt is automatically expanded into a full description
• Single-file weights come in bf16 / int8 (convrot) / nvfp4; the template defaults to int8 to save VRAM

Scope

• Good for: photorealistic portraits, product/commercial images, editing tasks that require changing part of an existing image
• Good for: multi-reference fusion (outfit swap, background swap, style transfer)
• Not for: anime specialization (this is a general-purpose photoreal-oriented model; for anime use the Anima / Illustrious family)

Tips

• When editing, describe what the result should look like, not "change red to blue" — the official template's example prompt is a full description of the finished image
• Reference images are downscaled to about 1MP; raise the reference resolution if you need detail (at the cost of VRAM and time)
• On a 16GB card, don't enable the full version, the prompt enhancer, and multiple reference images at the same time; stacking all three is the easiest way to run out of VRAM

Cautions

• Weights are in a gated repository; you must agree to the license on HF before downloading
• Raising CFG on the distilled version doesn't help — it only makes the image blurrier

Parameters

Full version

50 steps, CFG 3.5 (text-to-image) / CFG 1 (editing)

Distilled version

16 steps (4–16 possible), CFG 1

Resolution

Starting at 1024×1024; the official template uses ResolutionSelector to pick the aspect ratio

Text encoder

Qwen3-VL-4B (set CLIPLoader type to flux3)

VAE

FLUX.2 VAE (flux2-vae.safetensors)

Environment

ComfyUI 0.41.0 or later

Deployment

16GB verdict

Runs, using the int8 single-file path plus an int8 text encoder.

VRAM notes

int8 weights and the 4B text encoder must be loaded in stages on 16GB; the prompt enhancer is an additional 9B model (w4a8), so on 16GB it is best to turn it off first and enable it once the generation pipeline works.

Suggested starting point

Start with the Distilled 16-step version to confirm the environment, then switch to the full 50-step version to compare quality.

Comparison

Both from BFL: FLUX.2 [kotlin] has fallen to 430k downloads (last 30 days), while FLUX.3 klein reached 530k in ten days. The difference is that FLUX.3 merges text-to-image and instruction editing into one set of weights and adds a 4–16 step distilled version. The cost is a stricter license (FLUX Kommunity License) and, on 16GB, enabling the prompt enhancer adds a separate 9B model.