GH:42lux/ComfyUI-42lux-Hildegard-Refiner tile-based FLUX.2 Klein refiner
Nekodificator:
Other LoRAs:
ramonguthrie showed that qwen_3_vl_8b_nvfp4.safetensors works better than qwen_3_8b_fp4mixed.safetensors as TE for image editing with Klein
“Also you can use qwen_3_vl_8b_nvfp4.safetensors for this”
most people are making mistakes between Flux.2.0 Klein 9B and Flux.2.0 Klein 9B kv, if you are using Flux.2.0 Klein 9B kv [then use
Flux KV Cachenode]
victor.oli:
I made a multiple region masking, manual and automatic masking with sam3 using Claude Opus 4.8. It also has prompt enhancing with gemma4. If you want to test it GH:aiolicollective/aioli-node
2026.05
GH:ethanfel/Comfyui-PainterKleinImageEdit “node for FLUX.2 [klein] (4B / 9B) image editing and text-to-image generation”
GH:cosmicrealm/ComfyUI-Flux-FaceIR confusing tool for restoring faces on damaged photos.
David Shows frame by frame cleanup wf using Klein: david-show-klein-frame-by-frame-cleanup, david-show-klein-frame-by-frame-cleanup-2.
GH:supermansundies/comfyui-klein-edit-composite Klein edit composite node to mitigate color shift. Scf: “Sounds like you can do the same with masked inpainting in lanpaint”
Mark DK Berry on Kein:
took me months to realise it doesnt need speed upo loras. Klein changed my life. but license is meh.
4b - not good, 9b - good
if Klein is used for editing not creating images it seems the licensing becomes more of a grey area … 9b is non-commerical only
Klein can be used for editing
HF:wikeeyang/Flux2-Klein-9B-True-V2 a Klein finetune?
GH:Lakonik/ComfyUI-piFlow nodes for running AsymFLUX.2, a direct to pixels (PiD) flavor of Flux.2 dev?
Character trainer for flux: R:using_depth_maps_and_weight_noising_to_get_better
Flux Identity Adjuster V2 - node GH:Magirad/Flux_ID_Adjuster_V2, R:flux_identity_adjuster_v2
CA:2318844/klein-9b-body-weight-slider concept slider from thin to fat
Both base and distilled are distributed under a non-commerical license.
This is the faster model. Available in 4B and 9B, both distilled and base.
klein distilled is for converging in 8 steps. You might have augment the instruct with additional syntences to keep hair etc accurate.
This is the slow model.
Flux 2 is using Mistral as CLIP similar to z-image-turbo.
This is an 18Gb download named mistral_3_small_flux2_fp8.
bf16 version is 35.6Gb. Main model weights in fp8 are 35Gb.
amazing prompt adherence
flux2 fp8 it ran under 280 sec for 720x350 image [on RTX3060]
Flux 2 ran with fancy clownsampler can be slower than HunyuanImg3; at least HunyuanImg3 is just LLM speed
