KiwiMate Image Fast 1.0

A speed-tuned fine-tune of black-forest-labs/FLUX.2-klein-9B, trained with ai-toolkit to produce clean, sharp images in very few steps.

Image inputs supported — like its base model, this fine-tune can take images as inputs, so you can drive generation with reference photos, restyle or edit your own pictures, and do full image-to-image workflows, not just text-to-image.

What's in the repo

File	Description
`KiwiMate-Image-Fast-1.0-merged.safetensors`	Standalone merged checkpoint (~17 GB). Drop into `models/checkpoints` and run — no base model needed.
`KiwiMate-Image-Fast-1.0.safetensors`	Final LoRA adapter (step 3000, rank 32, ~158 MB). Apply on top of the original `flux-2-klein-9b` checkpoint at full strength.
`checkpoints/`	Intermediate LoRA checkpoints every 250 steps (250–2500), if you want an earlier snapshot.
`merge_lora.py`	Dependency-light script used to merge the LoRA into the base checkpoint (numpy only, streams tensor-by-tensor, resumable).

Sample gallery

All samples below were generated by the final step-3000 weights.



Prompt-driven scene: chess picnic with a fireball behind (color)	Monochrome café portrait, bokeh, holding coffee cup	Fisheye horse DJ at the decks under a disco ball	Beach guy giving thumbs-up next to a leaping shark

Cozy log cabin deep in snow-capped mountains	High-energy live rock singer on stage (B&W)	Craftsman hand-repairing a chair in his workshop (B&W)	Clean studio fashion shot on white seamless background

Text rendering: "thiss is a sign" held up in grainy monochrome	Sepia bulldog riding a chopper across the dunes	Neon KiwiMate storefront glowing on a rainy night	Bird soaring over misty peaks, ink-wash style (B&W)

Split-track commuter trains meeting head-on (B&W)

The set shows off speed-friendly fundamentals: good anatomy (hands, faces, animals), legible signage/text, strong lighting (stage rays, neon reflections), stylized monochrome looks, and larger scenes that hold detail without burning many sampling steps.

Usage

ComfyUI:

# Merged checkpoint — drag straight into models/checkpoints
huggingface-cli download KiwiMate/KiwiMate-Image-Fast-1.0 KiwiMate-Image-Fast-1.0-merged.safetensors --local-dir models/checkpoints

Or stack the adapter on the unmodified base instead:

# Standalone LoRA on top of black-forest-labs/FLUX.2-klein-9B at strength ~1.0
huggingface-cli download KiwiMate/KiwiMate-Image-Fast-1.0 KiwiMate-Image-Fast-1.0.safetensors --local-dir models/loras

For image-input workflows (reference images / edits / img2img), load your source picture the same way you would with stock FLUX.2 Klein — the LoRA is name-compatible with the native BFL single-file key layout (double_blocks / single_blocks).

Want to reproduce or tweak the merge itself? merge_lora.py rebuilds the standalone checkpoint from the base + LoRA pair:

pip install numpy
python merge_lora.py \
  --base flux-2-klein-9b.safetensors \
  --lora KiwiMate-Image-Fast-1.0.safetensors \
  --output merged.safetensors