FLUX.3 Image Klein (NVFP4) Tested: 5.3GB Model Runs Anime and Photoreal Scenes
Happy_Guy000 · reddit · 2026-10-16
A user tested FLUX.3 Image Klein (Base, NVFP4, 5.3 GB) in ComfyUI with the same workflow across three anime scenes and one natural-light still life to check style coverage.
Full setup:
- Models: flux3imagekleinnvfp4.safetensors; text encoder qwen3vl4bint8 (CLIPLoader type: flux3); flux2-vae
- Canvas: 1024x1344 portrait / 1344x768 wide, multiples of 16
- Sampler: euler + simple, 40 steps (official default 25; 50 slightly more detail); ersde worth trying
- CFG: 3.5 for anime, 2.5-3 for photo-style; empty negative prompt; no LoRAs, upscaler, or inpainting
- Node chain: UNETLoader > CLIPLoader(flux3) > CLIPTextEncode > EmptyFlux2LatentImage > KSampler > VAEDecode
Weights from black-forest-labs/flux-3-image-klein-comfy on Hugging Face (license acceptance required).
More from Multimodal
- MiniMax-H3 hailed as favorite open release of the year, runs fully local — AdinaYakup · 2026-10-23
- Michal Malewicz experiments with depth maps in his creative workflow — michalmalewicz · 2026-10-23
- One-Shot Tunable Style Transfer on Flux3 Via Reference Latent + Node — Winter_unmuted · 2026-10-23
- PhysVidBench accepted at EMNLP 2026: do video gen models grasp physical commonsense? — aykuterdemml · 2026-10-23
- EMNLP 2026 paper teaches multimodal LLMs to get cartoons via incongruity-resolution supervision — aykuterdemml · 2026-10-23
- A gorgeous Midjourney style: sref code 820221253 for dreamy lavender watercolor looks — umesh_ai · 2026-10-23