Fibo 2.0: Dev, Edit Turbo and Image Generation Costs

Dev is the full checkpoint for text-to-image, instruction editing and fine-tuning. Its reference recipe uses 50 steps and CFG 5. Edit Turbo distills the editing path to four steps without classifier-free guidance. Choose Dev when starting from text alone; Turbo requires an editing task.

Both use structured Visual GenAI Language (VGL) captions. A scene describes objects, lighting, composition and other attributes separately. That gives you explicit details to change between attempts, while the generated image still needs inspection for the requested words, object counts and preserved features.

Checked October 23, 2026, Bria lists Fibo generation and Fibo Edit at $0.032 per output image. fal’s bria/fibo-2/generate and bria/fibo-2/edit routes each quote $0.04 per image. At those rates, 1,000 output images cost $32 through Bria or $40 through fal.

Bria separately prices Fibo Structured Prompt at $0.02 per call. Generate can expand a short prompt into VGL internally; that does not mean adding a second Structured Prompt charge to every generation. Add the separate fee when your application makes that additional service call. A workflow making 1,000 such calls alongside 1,000 Bria image outputs would total $52.

Budget for attempts rather than just final assets. At fal’s rate, 100 generated outputs cost $4. If 25 are accepted, that generation spend is $0.16 per accepted image. Include any separately billed preparation, storage or editing stages in the same project budget.

Open the fal Generate playground, sign in and enter a scene description. Keep one output for the first attempt. For an edit, open the separate Edit route, attach your source image and state the change. That route accepts up to four reference images; assign a clear role to each one before combining them.

For a product edit, specify the background change and the features that should remain: label text, shape, materials and camera angle. After generation, download the result and inspect those details at delivery size. When several references are involved, check that texture, lettering and identity came from the intended source.

The 8B parameter count describes the diffusion transformer. Local inference also uses a Qwen3-VL-4B encoder and a FLUX.2 VAE. Include all components, runtime allocations and your chosen reference images when planning memory.

The Dev model card links Bria’s ComfyUI integration and a Diffusers quickstart. Its bria-vgl helper creates structured captions using Gemini, which needs its own API key and has its own service costs. Keep caption preparation separate from local image inference when comparing bills.

Comfy-Org places fibo2_dev_bf16.safetensors or fibo2_dev_int8_convrot.safetensors in models/diffusion_models, the matching qwen3vl_4b encoder in models/text_encoders and flux2-vae.safetensors in models/vae. Edit Turbo has its own diffusion files in the same directory.

In ComfyUI, start with a current Fibo 2.0 workflow and the matching diffusion, encoder and VAE files. Keep Dev at 50 steps and CFG 5 for the reference workload. Switch both the checkpoint and sampling recipe when testing Edit Turbo. Save the workflow, component revisions and output dimensions with each measurement.

Independent quantizer BennyDaBall reports these Dev results on an RTX 5090 with ComfyUI 0.42.3 plus native Fibo support, comfy-kitchen 0.2.39 and PyTorch 2.14 cu130. The graph uses Euler/simple, 50 steps, CFG 5 and 1024×1024 output. Times are medians of four to six warm executions with fresh seeds; whole-GPU peaks include the BF16 encoder, VAE and desktop.

The mixed NVFP4 result applies to that Blackwell setup. It does not establish a minimum-memory configuration or an RTX 4090 speed. The author’s separate Turbo repository reports sampling-plus-decode timings, which cannot replace Dev’s graph execution times in this table.

Bria releases the weights under CC BY-NC 4.0 for non-commercial use. Commercial deployment of those weights requires a separate agreement with Bria. Hosted service access has its own terms; the availability of a commercial API does not remove the self-hosting license condition.

For an occasional project, compare the per-image API bill with installation time and hardware costs. For repeated local work, divide the billed compute session, including setup and idle time, by accepted outputs. Keep Dev generation and Turbo editing as separate workloads.