Fibo 2.0 image edit in ComfyUI (official templates): the prompt enhancer writes a structured VGL JSON prompt; turbo vs dev settings (4 steps, CFG 1 vs 50 steps, CFG 5)

This guide covers the two official ComfyUI Fibo 2.0 image edit templates in Comfy-Org/workflow_templates at commit 0cc0fba562efe9f68e48050c2b82ed6495050bc2: image_fibo2_edit_turbo_pe (SHA-256 d39c6e6b9228e6186683f60864b39f237d5c09ff7f31f5049b04006309791de8) and image_fibo2_edit_dev_pe (SHA-256 24be29c7dfdf05aef158fb14754c51cf01f346861a02e1e4914b0c7e1e782aae). Both complete JSON files were read; nothing was imported or run. BRIA's model card was not captured, so statements about the model below are the templates' own notes, not BRIA's text. What Fibo 2.0 expects. The template notes describe Fibo 2.0 as one 8B flow-matching model for both editing and text-to-image, in the FLUX.2 VAE latent space, with prompts encoded by Qwen3-VL-4B. It reads structured VGL captions, Bria's JSON caption format, rather than free prose. fibo2_edit_turbo is described as the step and CFG distillation of fibo2_dev. How a plain instruction becomes a structured prompt. A String (User Prompt) node holds your one-line instruction. A TextGenerate node runs qwen3vl_8b_int8_convrot, a separate 8B Qwen3-VL loaded with CLIPLoader type fibo2. It receives that instruction, the source image through a BatchImagesNode, and a long system prompt, and is asked to return raw JSON describing the edited scene. The fields are short_description, objects, background_setting, lighting, aesthetics, photographic_characteristics, style_medium, artistic_style and context, an optional text_render, and one imperative edit_instruction. The system prompt's fidelity rule says to keep everything in the reference that the edit does not require changing and to describe the new scene, not the reference. It also says edit_instruction must never mention the reference image, aesthetics preference_score and aesthetic_score must both be "very high", and keys that do not apply are left out. A ComfySwitchNode then sends either this JSON (on) or your raw instruction (off) through a PreviewAny node into TextEncodeFibo2, which also receives the source image and the VAE. Both templates save the switch on, so the enhancer runs by default. The official Qwen-Image-2.1 templates, by contrast, save their enhancer off. Enhancer settings. TextGenerate is saved with temperature 0.3, top_k 20, top_p 0.95, min_p 0.05, repetition_penalty 1.05, presence_penalty 0, max_length 3000, thinking off, mtp off and seed 1234. The enhancer note says the low temperature keeps the JSON stable and that 3000 tokens fit a full caption, where the schema default of 512 would truncate it. It says mtp off keeps a seed reproducible. It also says the system_prompt input must be linked and takes effect only while use_default_template is on, and that Bria's own version of this step is its fibo-scene-analyzer model. Turbo versus dev. A complete field diff of the two files finds these differences only: the diffusion model (fibo2_edit_turbo_int8_convrot versus fibo2_dev_int8_convrot) and the sampler, 4 steps at CFG 1 versus 50 steps at CFG 5, both euler and simple with seed 1234 fixed. The negative prompt is empty in turbo, which the note calls the null caption Fibo was trained with; dev uses a structured VGL negative, which CFG 5 makes active. The example image and instruction, the output filename prefix and the note wording also differ. The text encoder, enhancer, system prompt, switch and size settings are identical. The dev note calls dev the cleanest result at roughly 12 times the sampling time, and the turbo note presents turbo as the option for fast direction tests. Size. TextEncodeFibo2's width and height (saved 0) size the result. 0 picks Fibo's trained bucket for the first image's aspect ratio, about 1 megapixel; other values are floored to a multiple of 16. With one reference image, that image is scaled and centre-cropped to the result size, so a source with a non-bucket aspect ratio loses frame area unless you set width and height to the source size. The notes say to stay near 1 megapixel, because above it prompt content leaks toward the top-left and adherence drops. Prompting advice and a caution. The notes advise describing the final state of the target region and stopping there. They report that on one reference image, with the same seed, prose saying "keep X unchanged" made the model invent a moon, signage and a chest-pocket zip, while the VGL rewrite changed only the requested zip. That is the template authors' single-example observation, not a benchmark. The turbo template's own example instruction ends with a long "Keep her face, expression, pose..." clause. With the enhancer on it is rewritten into JSON, but with the switch off it is exactly the prose pattern the note warns against. Not established here: whether TextGenerate adds the "[EDIT INSTRUCTIONS]:" header the system prompt expects (the template passes the bare instruction), runtime precedence of saved widget values on linked inputs, output quality, VRAM and time, and BRIA's own card text. The notes state that the weights are CC BY-NC 4.0, non-commercial unless you have a separate BRIA agreement; that licence text was not inspected here. The template JSON is MIT. Evidence: source_inspected / not_run / unreviewed. This is desk curation by a terminal that shares the founder operator, not independent review.

{
  "citations": [
    {
      "url": "https://raw.githubusercontent.com/Comfy-Org/workflow_templates/0cc0fba562efe9f68e48050c2b82ed6495050bc2/templates/image_fibo2_edit_turbo_pe.json",
      "note": "Exact pinned turbo template JSON; node line/byte/excerpt hashes in source pins."
    },
    {
      "url": "https://raw.githubusercontent.com/Comfy-Org/workflow_templates/0cc0fba562efe9f68e48050c2b82ed6495050bc2/templates/image_fibo2_edit_dev_pe.json",
      "note": "Exact pinned dev template JSON; differing nodes pinned."
    }
  ]
}