FIBO 2.0 Edit Turbo Does Text-to-Image Too: 4 Steps, No CFG, Tested on 22 Prompts
TimeTruth2490 · reddit · 2026-10-22
A Reddit user ran a deep dive (with sample images) on Bria's FIBO 2.0 Edit Turbo, showing the few-step, guidance-free editing model also does text-to-image:
- Two checkpoints: Fibo-2.0-dev (T2I + editing, 50 steps CFG 5, 100 forward passes) and Fibo-2.0-edit-turbo (4 steps, guidancescale=1.0, just 4 passes), sharing the same BriaFibo2Pipeline — pass no input image and it runs as T2I
- Structured prompts do the work: trained on Bria's VGL JSON captions, with a full Qwen3-VL-4B-Instruct as text encoder, the model can caption its own prompt using bria-vgl's open-source generate system prompt; plain prompts also work
- 22 prompts, same seed: self-written captions largely match typed prompts; object counts held (five friends, three kittens), text rendered correctly (a neon "Café Lumière" sign); captioning takes 70% as long as rendering
- Three failures: e.g. a top-down family dinner scene, each failing for a different reason
ComfyUI workflows and diffusers scripts for torch, MLX and mflux are on Hugging Face (lvladikov/FIBO2-Edit-Turbo-T2I).
More from Multimodal
- Video Arena Elo: Wan 3.0 Tops at 1173, Kling 3.0 Slumps to 17th — ArtificialAnlys · 2026-10-22
- Artificial Analysis Video Arena Turns Two: From 8 Models to 140+ Benchmarked — ArtificialAnlys · 2026-10-22
- Dev trains RAW Upscale MOSS detail LoRA with LoRA Dataset Studio, full recipe shared — Ill-Ant-9489 · 2026-10-22
- MiniMax H3 Singularity V2 updated, dual-sampling workflow released — theOliviaRossi · 2026-10-22
- MiniMax H3 Singularity V2 updated with dual-sampling workflow (duplicate) — theOliviaRossi · 2026-10-22
- AI-era composition is like photography: surf the collage manifold, don't paint — snikolov · 2026-10-22