I’m not sold on this one yet. I already expected no nudity handling, bummer i like the consistency it can bring in body shape,, but I’m also seeing issues with character consistency and fine details. For now, I still think Qwen 2.1 produces better character-training datasets. Check and curate the results carefully before training.

Promptwaffle WF — FLUX.3 Klein Base Character LoRA Pipeline V1

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

This ComfyUI workflow uses FLUX.3 Klein Base to generate portraits, expressions, body views, outfits, scenes, and lighting variations from your reference images. Four built-in prompt packs, automatic canvas selection, and a central control panel help you run a complete dataset without rebuilding the graph.

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

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

Features

• Three reference slots: face, full-body front, and optional full-body back, with individual switches.

• Four 70-prompt packs: Men, Women, Keep Identity, and Custom.

• Automatic canvas selection: square, portrait, or landscape based on the prompt row.

• Dual LoRA stacks: separate style/character and skin/detail stacks, with trigger-word controls.

• Base sampling: 50 steps, CFG 3.5, Euler, simple scheduler, and denoise 1.0.

• BF16 model and text encoder: the undistilled Base checkpoint and its required FLUX.3 text encoder.

• Approximately 2 MP output: with reference images fitted to approximately 1 MP.

• Optional PW Color Tools: palette extraction, grading, curves, optics, grain, and look/LUT export.

• Organized output: character-based naming, PNG metadata, and raw-versus-graded comparison.

• Memory controls: optional text-encoder unloading after encoding.

The Keep Identity pack asks the model to preserve clothing and hairstyle. The Men and Women packs introduce more variation. Reference images guide identity and appearance, but consistency is not guaranteed. Check and curate the generated images before training.

Required model files

Paths below are relative to your ComfyUI installation.

ComponentExact filenameFolderBase diffusion modelflux3_image_klein_bf16.safetensorsmodels/diffusion_models/Text encoderqwen3vl_4b_bf16.safetensorsmodels/text_encoders/VAEflux2-vae.safetensorsmodels/vae/

Direct downloads:

• FLUX.3 Klein Base BF16 diffusion model

• Qwen3-VL 4B BF16 text encoder for FLUX.3

• FLUX.2 VAE

• Official repository, access terms, and license

Sign in to Hugging Face and accept the repository’s access terms before downloading.

Use the Base checkpoint listed above. Files containing _sd_ are distilled alternatives with different sampling defaults.

Although the text encoder is named Qwen3-VL, it is the encoder required by FLUX.3. Set its loader type to flux3, not qwen_image or flux2.

No additional LoRA or prompt upsampler is required. Both LoRA stacks start empty. Only load adapters explicitly compatible with FLUX.3 Klein Base. Qwen-Image, FLUX.1, and FLUX.2 LoRAs are not interchangeable with this model.

Required custom nodes

Install these packs through ComfyUI Manager. If a pack isn’t listed, use its link below.

• LoRA Manager — Loads LoRAs, manages trigger words, and saves image metadata.

• Switchboard — Turns workflow sections on or off.

• rgthree — Routes the selected prompts and images.

• Impact Pack — Chooses the canvas orientation for each prompt row.

• Comfyroll — Runs the built-in prompt lists.

• Resolution Master — Controls output dimensions.

• PW Color Tools — Adds optional colour grading, curves, optics, and grain.

• Deno custom nodes — Provides optional text-encoder unloading to free memory.

Install all eight packs, including those used by optional sections, so the supplied workflow loads completely.

Installation

1. Install through Manager, or use the links above.

2. For manual installs, place each pack’s folder in ComfyUI/custom_nodes/ and follow its README.

3. Restart ComfyUI, then press Ctrl+Shift+R in your browser.

ComfyUI version

Use ComfyUI 0.41.0 or newer with local FLUX.3 support.

ReferenceMultiLatent and ImageCompare are built-in ComfyUI nodes. You don’t download them separately. If ReferenceMultiLatent appears as an “unknown pack,” update ComfyUI core.

In the text-encoder loader, select flux3 as the type.

ComfyUI version

Use ComfyUI 0.41.0 or newer with local FLUX.3 support.

ReferenceMultiLatent and ImageCompare are built-in ComfyUI nodes. You don’t download them separately. If ReferenceMultiLatent appears as an “unknown pack,” update ComfyUI core. The text-encoder loader must also offer flux3 as its type.

Quick start

1. Load PW_FLUX3_Klein_Base_CharacterLora_WF_V1.json.

2. Replace the saved reference images with your own character images.

3. Enable the reference slots you want in the REFERENCE PANEL.

4. Set OUTPUT FOLDER and CHARACTER / FILE NAME.

5. Enable exactly one prompt pack in the Control Panel.

6. Leave 0c reference canvas OFF to use the saved 2 MP output sizes.

7. Reset PROMPT INDEX to 0.

8. Generate one test image and check it before queuing a full batch.

9. Reset the index to 0, then queue 70 runs for the complete pack.

Each queued browser run uses one prompt row. The index increments automatically and must be reset before starting another batch.

Canvas sizes

Prompt rowsOrientationSaved output size1–18Square1408 × 140819–64Portrait1280 × 160065–70Landscape1600 × 1280

For custom lists, keep this row order or adjust SQUARE ROWS END and VERTICAL ROWS END.

Each enabled reference image is fitted to approximately 1 MP. Enabling 0c reference canvas overrides the sizes above and uses the first enabled reference’s fitted dimensions.

Notes

• This is a single Base-model workflow. There is no Turbo mode.

• Sampling defaults are 50 steps, CFG 3.5, Euler/simple, denoise 1.0. Image quality still depends on the prompt, references, and seed.

• The negative caption is empty by default. Enable USE NEGATIVE PACK to use the selected pack’s negative prompts. At CFG 3.5, they can affect the result.

• Reference images are applied to both positive and negative conditioning.

• Describe the complete target image, including framing, clothing, pose, background, and lighting.

• Reference labels start at <ref_image_0>. Disabling a reference compacts the numbering.

• Leave colour grading off for clean dataset generation, or configure your own grade first.

• If you encounter out-of-memory errors, try the Unload Text Encoder switch or a smaller output canvas.

• The Men and Women packs include nude adult body-reference prompts in rows 42–49. Remove those rows if unwanted. The Keep Identity pack excludes them.

• My concerns about nudity handling and dataset quality are my current assessment, not a controlled benchmark.

• Reference images and model weights must be supplied separately.

Built by Promptwaffle

More workflows, LoRAs, and models on Civitai · GitHub