Open-source style transfer methods compared: 2026 update tests USO, Kontext, Krea2 without Loras
Dry-Resist-4426 · reddit · 2026-09-14
The author updates their September 2025 comparison of state-of-the-art open-source style transfer methods — turning an image into the style of a reference image while preserving content, ideally with no Lora training and no hand-written style prompts.
- Context: ByteDance's USO release (Aug 2025) prompted re-evaluation against existing methods.
- Setup: ForgeUI (SD1.5, SDXL with two CLIP variants) and ComfyUI 0.34.0, uniform 1024x1024, canny ControlNet for consistency in most runs.
- Prompting: basic captions; Kontext uses a maintain-everything prompt; Krea2 uses a CLIP+generate-text node to auto-write a style-only description (full prompt-engineering prompt included), combined with a restyle instruction. Krea2's Depth Lora gave no improvement and was excluded.
- Results involve mild cherry-picking (best of 3-5); workflows are shared for reproducibility.
More from Multimodal
- Dissecting 'The Gen Alpha Melody': what AI-generated pop songs share and where they differ — caglar_ee · 2026-09-14
- AI-generated dancer Astra gets stage upgrade as limb-merging artifacts improve — jacalulu · 2026-09-14
- This Gemini prompt turns your photo into an 80s Indian action movie scene — HeyAmit_ · 2026-09-14
- Blender-to-Seedance 2.5 workflow demo turns previews into final AI video — azed_ai · 2026-09-14
- BOTTOM GUN parody made with the new Flux Video Edit model and video continuation — MoistRecognition69 · 2026-09-14
- H3 text-to-video recreates a century of film eras with generic prompts at int8/20 steps — SIR_NVAX_A_LOT · 2026-09-14