Promptwaffle WF — FIBO 2.0 Character LoRA Pipeline V2

One character. Every angle.

Turn your character references into a varied image dataset for your next character LoRA.

This ComfyUI workflow brings together FIBO 2.0 Edit Turbo and Dev, four expanded prompt packs, structured JSON captions, automatic canvas selection, and a central control panel.

Use Edit Turbo to build variations from your character references, or switch to Dev for standalone text-to-image generation. Create portraits, expressions, body views, poses, outfits, scenes, and lighting variations without rebuilding the graph.

This workflow generates dataset images. LoRA training and dataset captioning are handled separately.

Need help organizing and tagging your dataset? Try my Idiot LoRA Builder.

Features

Two generation modes: reference-based Edit Turbo and standalone Dev text-to-image.
Three reference slots: face, full-body front, and optional full-body back, with individual switches.
Four 110-prompt packs: Men, Women, Keep Identity, and a preloaded Custom pack.
Structured JSON captions: prepared target descriptions and editing instructions for FIBO.
Final prompt preview: inspect the exact positive JSON sent to the encoder.
Automatic canvas selection: square, portrait, or landscape based on the prompt row.
Dual optional LoRA stacks: separate character/style and detail stacks, with trigger-word controls.
Native output above 2 MP: no final upscale required for the saved canvas sizes.
Optional PW Color Tools: palette extraction, grading, curves, optics, grain, and look/LUT tools.
Organized output: character-based naming, PNG workflow metadata, and raw-versus-final comparison.
Memory controls: optional text-encoder unloading and a smaller trained-size Turbo canvas.

The prompt packs cover close-ups, expressions, gaze direction, profile views, standing and seated poses, hands, walking, lighting, and wider compositions.

Reference images guide identity and proportions; they do not lock them. Review and curate your generated images before training.

Edit Turbo versus Dev

SettingEdit TurboDevMODE switchONOFFInputEnabled references and selected prompt packSeparate dev_prompt in P2Saved steps430CFG15Image referencesUsedBypassedPrompt-pack behaviorAdvances through the selected positive rowsRepeats the standalone Dev caption

Both modes use Euler, the simple scheduler, and denoise 1.0.

The 110-row reference packs are designed for Edit Turbo. Dev does not automatically turn those rows into 110 different text-to-image prompts.

Required model files

Paths below are relative to your ComfyUI installation.

ComponentExact filenameFolderEdit Turbo diffusion modelfibo2_edit_turbo_int8_convrot.safetensorsmodels/diffusion_models/Fibo2/Dev diffusion modelfibo2_dev_int8_convrot.safetensorsmodels/diffusion_models/Fibo2/Text encoderqwen3vl_4b_int8_convrot.safetensorsmodels/text_encoders/VAEflux2-vae.safetensorsmodels/vae/

Model downloads and documentation:

In the text-encoder loader, select fibo2 as the type.

No additional LoRA or external vision-language model is required. Both LoRA stacks start empty. Only load adapters confirmed compatible with FIBO 2.0 and your selected mode.

The structured-prompt helper formats prepared captions. It does not inspect or automatically caption your reference images.

Required custom nodes

Install the following packs through ComfyUI Manager, or use their repositories:

LoRA Manager — LoRA loading, trigger-word controls, and image metadata.
Switchboard — Workflow switches and mode routing.
rgthree — Group controls and routing.
Impact Pack — Prompt-index comparisons and canvas routing.
Comfyroll — Built-in prompt lists.
Resolution Master — Canvas dimensions.
ComfyUI Essentials — Supporting math and utility nodes.
PW Color Tools — Optional grading, curves, optics, and grain.
Deno custom nodes — Optional text-encoder unloading.

Installation

Install the custom-node packs above.
Put the model files in the folders listed above.
Update ComfyUI to a build with native FIBO 2.0 support.
Restart ComfyUI, then press Ctrl+Shift+R in your browser.
Load PW_Fibo2.0_CharacterLora_WF_V2.json.

The PW helper requires no additional pip dependencies.

ComfyUI support

Your ComfyUI installation must include TextEncodeFibo2, the fibo2 text-encoder type, and EmptyFlux2LatentImage.

If those nodes or loader options are missing, update ComfyUI core. Installing custom-node packs alone will not add native FIBO support.

Quick start — Edit Turbo

Load PW_Fibo2.0_CharacterLora_WF_V2.json.
Select the required model files.
Leave MODE ON for Edit Turbo.
Replace the saved reference selections with your own character images.
Enable REF 1 for the face and REF 2 for a matching body view. Enable REF 3 only if needed.
Set OUTPUT FOLDER and CHARACTER / FILE NAME.
Enable exactly one prompt pack in the Control Panel.
Leave 0b automatic canvas ON and 0c trained reference size OFF for the saved output sizes.
Reset PROMPT INDEX to 0.
Inspect P3 FINAL FIBO JSON, then generate one test image.
Review the pack’s content and your test result before batching.
Reset the index to 0, then queue 110 single-image runs for a complete, unmodified pack.

The index increments automatically and does not wrap. Reset it before starting another pack.

Quick start — Dev

Set MODE OFF.
Edit the separate dev_prompt input in P2.
Enter a standalone description or structured caption JSON.
Keep one prompt pack enabled for routing.
Choose your canvas and generate a test image.

Dev bypasses the reference images in this graph. Its standalone caption repeats while the prompt index and automatic canvas advance.

Canvas sizes

Prompt rowsZero-based indicesOrientationSaved output size1–260–25Square1440 × 144027–9626–95Portrait1280 × 192097–11096–109Landscape1920 × 1280

These are native generation dimensions, approximately 2.07–2.46 MP.

For custom lists, maintain this ordering or adjust SQUARE ROWS END and VERTICAL ROWS END, saved as 26 and 96.

0b ON: canvas orientation follows the prompt index.
0b OFF: every row uses the portrait/one-size canvas.
0c ON in Turbo: overrides the normal canvas with a trained size around 1 MP, based on the first enabled reference’s aspect ratio.
Dev ignores 0c.

The 0c option does not preserve the reference’s exact pixel dimensions.

Working with structured prompts

Each positive pack row is one complete JSON object on one physical line.

Edit the caption fields inside that object. Keep the CR Prompt List prepend and append fields empty, and do not add prose after the closing brace.

The helper handles JSON serialization and merges enabled LoRA trigger words into artistic_style. P3 previews the resulting positive caption.

Prepared structured captions are included, but this is not automatic image captioning or a vision prompt upsampler.

Prompt-pack content

Review the rows before queuing a full batch.

Men, Women, and Custom: include adult nude body-reference prompts in rows 50–57.
Every pack, including Keep Identity: includes non-sexual adult nude anatomical studies in rows 85–96.
Those anatomical rows specify clearly adult references age 25 or older.
Keep Identity preserves clothing elsewhere, but it is not an entirely clothed pack.

Remove or replace unwanted rows before batching. Keep positive and negative lists aligned, and adjust the canvas boundaries if deleting rows changes their positions.

Notes and tested limits

Turbo uses an empty negative caption at CFG 1. Dev uses the selected pack’s paired negative row.
Reference markers use <image_1>, <image_2>, and so on. Disabling references changes their numbering.
Keep REF 1 enabled for the supplied face-anchor instructions. A clear, single-person body image works best for REF 2.
Multi-view reference sheets can produce duplicated subjects. Inspect outputs carefully.
Body-proportion instructions are prompt guidance, not a measurement or shape lock.
Colour grading and diagnostic stage saves start OFF. Configure and test a grade before applying it to a dataset.
The comparison viewer shows raw decode versus final output, not Turbo versus Dev.
For memory pressure, try batch size 1, fewer references, encoder unloading, a smaller canvas, or Turbo’s 0c option.
The supplied workflow notes report successful RTX 4090 24 GB tests in both modes. Exact framing remained imperfect.
All 440 positive captions and their paired negatives passed structural checks, but the full expanded packs have not been visually validated.
Optional LoRAs, grading, and anatomical rows were not covered by the reported visual tests.
Reference images and model weights must be supplied separately.

Model license

BRIA provides the model weights for non-commercial use under its published license. Review the current terms and obtain the appropriate BRIA agreement for commercial use. Other assets and LoRAs may have separate terms.

Built by Promptwaffle