Character Dataset Creation — Qwen Image 2.1 Sheets

Build a recurring character from a face, your chosen body shape and a wardrobe reference, approve a front/back/portrait sheet, then generate individual images with varied framing, poses and expressions.

The workflow is downloaded from this Civitai resource. Eclipse supplies the custom nodes and editable prompt defaults.

Original workflow credit: This workflow builds on v1 of Qwen Image 2.1 Character Reference Sheet Generator — Face, Wardrobe. Credit goes to its original creator for the base workflow and face/wardrobe reference-sheet approach. This adaptation extends it with physique references, generated mannequin guides, sheet splitting, pose planning and varied character dataset generation.

What this workflow does

Uses an existing face, extracts a headshot, or generates an identity from text.
Builds a wardrobe reference board from an outfit image.
Generates a three-view mannequin guide using body presets, with an optional Reference mode for physique images.
Combines mannequin proportions, facial identity and wardrobe into one character sheet.
Crops that sheet into separate front, rear and portrait references.
Plans 50 individual dataset images at the saved count, with three prompt candidates per shot and one selected image per shot.
Includes a separate Flux 2 edit branch, image tagging and caption export.

The workflow has nine groups: six character/dataset stages followed by Flux2 Edit, Image to Prompt and Save Prompts.

Face source or generated face ──────────────────┐
Outfit source → wardrobe reference board ───────┼→ Character sheet
Body presets / references → mannequin guide ────┘        ↓
                                               Front / rear / portrait crops
                                                        ↓
                                               Shot planner → individual images
                                                        ↓
                                               Optional Flux edits and captions

Before the first queue

Install a ComfyUI build with Qwen Image 2.1 support and the following packages, then restart ComfyUI:

ComfyUI Eclipse: Reference preparation, sheet packing/splitting, shot planning, Set/Get routing, review previews, sampling, folders, saving and prompt utilities.
ComfyUI Smart Model Loader: Model/CLIP/VAE loading, pipeline outputs and the Qwen Image 2.1 encoder adapter; also the Flux edit loader.
ComfyUI SmartLLM: Smart LM Loader for the Image to Prompt branch. Needed when using that branch.

Use an Eclipse installation containing Character Reference Prepare/Select, Character Sheet Pack/Split and Character Shot Planner. This graph does not contain the old Smart Detection, Krea 2 or KJNodes stages.

Main model selections

The saved main loader selects these local files, in folders under ComfyUI/models/:

diffusion_models: qwen_image_2.1_int8_convrot.safetensors
text_encoders: The saved filename is qwen3-vl-8b-heretic-1.3.0_int8_convrot.safetensors. The Hugging Face INT8 ConvRot build is named qwen3-vl-8b-heretic-1.3.0-int8convrot.safetensors; select the downloaded filename in the loader.
vae: qwen_image_2.1_vae_bf16.safetensors

Models are not included in the workflow. Select the actual installed filenames. These selections describe this saved setup, not a requirement to use that exact quantization; substitutes must be compatible with Qwen Image 2.1 and your installed loader.

Choose which optional stages to run. For the first character review, mute Flux2 Edit, Image to Prompt and Save Prompts. Enable them later after setting your edit instruction and caption preferences. Review stops let you approve intermediate results before continuing.

Choose your own image in every source loader you intend to use.

01 · Models, project and shared specifications

The main Smart Model Loader supplies the model, encoder, VAE and sampling settings through the shared pipe_main channel. The model cache output travels through model.

Edit the project/character name, dataset scene, identity details, body proportions and wardrobe specification here. The character name controls both the planner project history and filename subfolder. Use a new name for a new character or fresh reservation history.

The current effective settings are:

Main sampler: Euler, normal scheduler, 30 steps, CFG 3.5, denoise 1
Dataset canvas: 768 × 1376 from Smart Folder
Landscape sheet dimensions: Shared height × width: 1376 × 768 with the settings above
Qwen encoder resolution: 0; use the encoder's latent output
Planner: 50 shots, balanced selection, any body mode, all pose categories, preview operation
Planner cooldowns: Pose 3, expression 3, camera family 1
Sheet splitter: Auto, gutter 0.004

Use the main loader, Smart Folder and Pack report to check the effective configuration. At CFG 1, negative conditioning is inactive; CFG 3.5 uses it.

Face, wardrobe, mannequin and assembly have separate Seed nodes. The dataset seed feeds the planner, which supplies an aligned seed for each chosen shot. Keep the relevant seeds fixed while comparing prompt changes; changing the planner seed can allocate a new reserved batch variant.

02 · Face identity

The saved face mode is extract. Character Reference Prepare supplies prompts and the source image; Character Reference Select chooses the appropriate result.

reference: Use the supplied face image directly; skip preparation generation.
extract: Generate an isolated headshot from the source (awesome for lowres images!).
generate: Generate from the written prompt without using the source image.
disabled: Omit the reference. This is not valid for the later sheet stage, which requires a face.

Edit the connected FACE positive and negative text nodes. When switching from extraction to generation, also change the positive so it describes a face rather than referring to a missing source image. Review identity, hair and skin appearance before moving on. Use the upstream resize/crop controls when necessary.

The face is the identity reference.

03 · Wardrobe reference board

Load the outfit source and edit the wardrobe specification. The board can show front, back, side and detail views of one outfit. Repeating an item in another view is different from adding another physical garment.

The prompt preserves construction, layers, colors, materials, closures, straps, accessories and footwear. For a cropped source, use the wardrobe text to specify missing lower-body clothing or shoes. Those additions are design choices, not recovered information from outside the source frame.

Check the board before assembly: missing items, altered fastenings, extra garments or wrong colors can propagate through the whole dataset. Avoid conflicting outfit descriptions in the shared text.

To skip the board, disable use_wardrobe on the SHEET Pack, provide a complete written outfit specification through that Pack's details input, and mute the wardrobe group if you also want to skip its independent preview. To reuse a previously approved board, load it directly into the SHEET Pack's wardrobe input.

04 · Choose and review the body shape

The mannequin is a shape guide, so detailed skin, polished shading and perfect hands are not the priority. Look for the body proportions you want in all three panels: full-body front, full-body rear and an enlarged upper-body portrait. No mannequin template is required.

Start with Preset

The saved workflow uses Preset mode. Choose Gender, then adjust Build, Muscularity, Breast size, Chest breadth, Hip width and Buttock size below Stage.

Unspecified leaves that attribute open; it does not mean medium or zero.
Body controls are available for every Gender. The model can still associate gender with particular proportions, so review the result.
Use Body proportions for a short clarification of a particular feature. Keep it consistent with your selections.
Preset ignores reference images. You can choose the shape entirely through these controls and text.

The saved selections are examples, not a neutral starting body. Choose your own settings before generating. To start broadly neutral, choose average Build and leave the other body attributes unspecified.

Faster shape previews

The saved workflow uses 30 steps from the shared loader. To try the faster setting only for mannequins, disconnect the steps connection on the mannequin sampler and set its steps to 20. Keep the other sampler connections. Changing steps in the main loader would also affect other stages.

What to check before approving

The presets guide the model; they are not exact measurements. Across 35 test renders:

Slim/heavy build, high muscularity, very large breasts: Changes were clearer, but could also change nearby regions.
Flat or very small breasts: Often remained rounded or raised.
Hip width and chest breadth: Narrow and broad settings sometimes looked too similar. Stronger wording could create unnatural bulges.
Buttock size: Differences were weak; the fixed front/rear views also make projection difficult to judge.

Check the front, rear and portrait for the same physique. Try one change at a time with a fixed seed, then check another seed before settling on wording. A longer or shorter prompt is not automatically better.

Use physique images when needed

Choose Reference to guide the physique with front and/or rear images. This saved workflow has no body-source branches connected: add a Load Image node for each source you want, connect it to the Pack's front_body or rear_body input, and enable the matching use_front_body or use_rear_body switch.

Reference mode ignores the hidden body presets. Gender and written details still apply. Both photographs and illustrations can guide visible shape, but loose clothing cannot reveal hidden anatomy. An optional layout image guides framing; leaving it disconnected uses the written three-panel layout. Missing a rear source does not remove the rear output panel.

Body references are used only for the mannequin stage. The later character sheet uses the approved mannequin, face and wardrobe; raw physique images do not pass directly into the dataset.

REVIEW MANNEQUINS has Stop enabled in the saved file. Approve the shape and all three views, then turn Stop off and queue again to continue.

05 · Assemble and approve three character panels

Reference order is deliberate:

<image1> — mannequin guide: Body proportions, pose, panel layout, subject scale and portrait framing.
<image2> — face: Facial identity, hair and skin appearance.
<image3> — wardrobe board: Garment design, colors, materials, layers, accessories and footwear.

The assembly prompt asks the model to turn the guide into photographs of one living person, retaining chest volume, musculature, hips and glutes while replacing clay with human skin and fitting the outfit to that body.

Review all three views for identity, physique, outfit completeness and consistent garment placement. Stronger prompts improve control but do not guarantee exact body geometry, muscle definition, footwear or small garment details.

The saved REVIEW CHARACTER SHEET Stop is off; enable it for your first review. The subsequent PORTRAIT crop Stop is on.

Split the sheet before generating variations

Character Sheet Split produces three separate images, not an image batch. Choose auto to divide the actual input image width into three equal parts on every run, including after a resolution change. No width input is needed.

The workflow is saved in auto mode. Choose manual for uneven panel layouts and adjust front_end and rear_end using the previews. Mode is the top widget: auto hides these two boundary controls, and manual restores them with their saved values.

Gutter trims the left and right edges of each panel to remove divider lines; it does not add spacing. It stays visible in both modes and is a fraction of the full sheet width. The saved 0.004 trims about 6 pixels from each edge at 1376 pixels wide. Set it to 0 for complete thirds without trimming. Keep clothing and body outlines inside the remaining crop.

The dataset uses the front crop, portrait crop and optional rear crop. If the portrait loses identity, connect the reviewed original face to the dataset Pack's face input instead. Turn off the portrait review Stop when all crops are approved.

Separate Save Images nodes save the complete sheet and each of the three crops, using the prefixes sheet, frontal, rear and portrait.

06 · Varied individual dataset images

The dataset Pack sends the approved front character, a separate face reference and optional rear character to Qwen. The full sheet, raw body sources and original wardrobe board do not enter this stage.

rear_body and rear are different: the first is the original physique photograph used to build mannequins; the second is the finished rear character crop used for continuity. A rear reference does not request a rear-facing output. The stock planner varies front, three-quarter and profile orientations; adding rear conditioning does not add rear shots to its camera pool.

How the shot planner works

For each requested shot, the planner constructs close, mid and wide candidates with distinct poses and expressions. It selects one according to balanced, a fixed distance, or your choices sequence. 50 selected shots generate 50 images, not 150. Keep one latent image per planned prompt.

body_mode selects seated, standing, lying, or any posture. Seated includes varied benches, stools, sofas, floor cushions and supported reclining.

pose_category independently selects everyday, sports, or all activities. For boxing, martial arts stances and kicks, use standing + sports + wide selection. Close shots show athletic preparation instead; seated/lying sports use stretches and recovery poses. These filters do not change the outfit or location. Edit the pose descriptions and their optional category in prompts/shot_planner/poses.json; camera wording for lying poses is in camera.json.body_overrides.lying. Use a new batch ID after editing files so a reserved batch does not replay its old prompts.

Start with count=3 and operation=preview. Preview does not reserve camera combinations, but its outputs can still generate images when the downstream sampler is active. To inspect text only, execute the planner/report branch without generation.

When satisfied, choose reserve for persistent history. The ledger records all three camera candidates, including the two not selected. Exact camera combinations stay unavailable in that project; pose, expression and camera-family cooldowns are temporary.

Reuse batch ID and matching settings: Replay the stored plan and seeds.
Change planner settings under the same base ID: Resolve a numbered variant such as sheet-001-2; the report shows the actual ID.
Use a new batch ID: Create fresh shots within the same project's remaining pool.
Change project name: Start separate history.
Retry after failed/cancelled image generation: Reuse the reserved plan; failed images do not release reservations.

A batch ID is not a reset button. Ledgers are stored under output/eclipse/shot_ledgers/; the report gives the exact path. For testing, use preview or a separate test project. To reset the same project deliberately, stop using it in running/queued jobs and move its ledger to a backup location before reserving again. Old batch replay is then unavailable until that ledger is restored.

count accepts 1–100 selected images per run. With the current default camera pools, a fresh project has 225 close, 270 mid and 225 wide combinations. Every selected shot consumes one candidate of each distance, so 225 selected images per project is an upper bound, not a guaranteed capacity. Cooldowns, filters and previous reservations can reduce the usable total. Preview does not consume that capacity.

Reservations are atomic: either the complete plan is saved or none of it is. Smaller batches and lower cooldowns can use more of the remaining pool. Changing only the batch ID does not replenish used camera combinations.

Customize the camera, pose and expression examples

Edit these files inside Eclipse, then click Reload planner files:

prompts/shot_planner/camera.json
prompts/shot_planner/poses.json
prompts/shot_planner/expressions.json
prompts/shot_planner/rules.json

Keep compatibility rules meaningful: walking needs visible legs, and table-height framing belongs in close/mid shots. New options need stable IDs. Reserved batches retain their saved pools; use a new batch ID to apply changed pool files.

The planner is a prompt planner, not an image-quality checker. It does not detect near-duplicate output images or reject malformed anatomy automatically. Review and curate the generated dataset.

Editing positive and negative prompts

Face, wardrobe, character-sheet and dataset stages have connected prompt text boxes. Edit those for the active workflow. The mannequin uses its editable file defaults instead; its positive and negative overrides are disconnected. Connected overrides replace the corresponding file text; editing a default JSON alone does not update saved multiline overrides.

Face Prepare defaults: prompts/character_references.json.
Wardrobe, mannequin, sheet and dataset Pack defaults: prompts/character_sheets.json.
Disconnect an override to use its file default. A connected empty negative intentionally clears it.
Face Prepare's positive override replaces its description/instructions. Sheet Pack still supplies resolved reference roles and details; the mannequin also adds its fixed three-panel layout, Gender and active Preset selections. Written details refine the attributes you name.
Keep role names such as LAYOUT GUIDE, FACE, WARDROBE, FRONT BODY and REAR BODY in Pack prompts. The node resolves them to the actual <imageN> numbers, including when optional references are omitted.
Connect the Pack's positive output to Show Any to inspect what reaches the encoder.

Prompt-only file changes do not need a server restart. Reload the downloaded workflow to pick up changed saved text, or copy the changed prompts into your current graph after preserving unsaved edits.

Write direct instructions describing the intended image, the changes to make and the features to preserve. Use the Pack's uppercase role names for references and let it assign the image numbers. For mannequin prompts, use plain prose instead of a page of headings or labels. Do not add chat-template or vision tokens manually; the encoder supplies its own model formatting.

Use external prompt files instead of overrides

Disconnect positive_override and/or negative_override on the relevant Prepare or Sheet Pack node to use file-based prompts. Each input works independently: you can keep a positive override connected while loading the negative from a file, or the reverse. Leaving a connected text node empty does not enable the file default.

Open the following files in a text editor, relative to your Eclipse installation (ComfyUI/custom_nodes/comfyui_eclipse/):

prompts/character_references.json: For Face Prepare: prepare → face → extract or generate for the positive, and prepare → face → negative for its negative. Choose the positive matching the node's mode.
prompts/character_sheets.json: Stage prompts: wardrobe, sheet, dataset, with matching _negative entries. Mannequin Preset uses mannequin_preset; Reference uses mannequin. Both use mannequin_negative and mannequin_layout. Selection wording is under mannequin_attributes; mannequin_details introduces your written clarification.

Back up the file, edit the text values, save, and queue again. The nodes reload file defaults automatically; no restart or planner reload button is needed. Keep valid JSON: double-quoted strings, \n for line breaks inside strings, escaped quotes (\"), and no comments or trailing commas. Keep schema_version unchanged and retain the uppercase reference-role names used by Sheet Pack.

Edit the runtime files under prompts/, not the distributed .defaults/ examples. These are shared defaults, so changes affect every graph using that file without the corresponding override. The face/face_negative entries in character_sheets.json do not control Face Prepare; it reads character_references.json. Connect the node's positive/negative outputs to Show Any to verify the resulting text. For dataset changes, check the planner report as well: an existing reserved batch replays its saved shot prompts.

Flux2 Edit — separate optional branch

This branch collects only img_varied: the individual dataset images. With 50 variations, it processes 50 edit sources. The complete sheet and reference crops have their own savers and are not included in this edit collection.

The saved edit setup selects:

Diffusion model: Dark Beast — saved Civitai version, saved locally as darkBeastMar0326Latest_dbkleinv2BFS.safetensors. This version now requires payment.
Text encoder: qwen_3_8b_fp8mixed_abliterated.safetensors
VAE: flux2-vae.safetensors
Loader sampler settings: Euler/simple, 5 steps, CFG 1

Alternative: Use standard FLUX.2 Klein 9B plus a BFS (Best Face Swap) LoRA. and or a NSFW LoRa. Select a flux-klein_9b LoRA, place it in ComfyUI/models/loras/, and apply it to the edit model before sampling. BFS also offers versions for other model families; always choose the LoRA matching your model. Follow the BFS author's guide for that version's reference order, prompts and LoRA settings.

These files belong in the usual diffusion-model, text-encoder and VAE folders. They are only needed for this branch. The embedded subgraph supplies reference conditioning and scheduler selection; Fast Mode Toggle/Mode Bridge controls the scheduler route and review Stop. Use the review path before allowing the edit saver to continue.

Edit the branch's text input to describe your intended transformation. For a correction pass that preserves the character and outfit, an example is:

Preserve the same identity, body proportions, outfit, pose, expression and
framing. Correct malformed details and visible artifacts while retaining
natural skin and fabric texture. Keep the existing garment design and colors.

For a complete run with 50 variations, the savers produce 4 reference images + 50 base images + 50 edited images = 104 images. Without Flux edits, that is 54 images.

Image to Prompt and Save Prompts

The saved tagger is SmartLLM's WD14-eva02-large-v3 (SmilingWolf/wd-eva02-large-tagger-v3 on Hugging Face), with general threshold 0.35, character threshold 0.85 and underscore replacement. Its output passes through Filter Prompt and a review stop before export. Inspect the filter: it removes several appearance and style tags, which may not suit your character. Other vision-capable caption models require their own compatible configuration.

The caption collector takes img_varied, then img_f2edit. The path collector matches that order with files_varied, then files_f2edit. At 50 variations, this gives 100 captions when both routes complete, or 50 without Flux edits. The four reference images are saved separately and are not part of caption export.

Keep image and path collections in the same order when customizing these stages. To caption additional saved images, add their image and file-path channels together.

The export settings are TXT, %source_filename, keep mode, and %source_base_folder/prompts. Existing captions are retained. Change the folder expression to %source_base_folder if your training tool expects sidecars beside the images. Switch to overwrite only when intentionally replacing reviewed captions.

Output folders

Smart Folder supplies the output path through Set/Get. The saved root is images/Datasets, with a date subfolder enabled. Add Filename Prefix adds your character name as a subfolder; base files use image, edit files use flux2.

A typical completed run is organized like this:

ComfyUI/output/images/Datasets/<date>/<character-name>/
├── sheet_0001.png
├── frontal_0001.png
├── rear_0001.png
├── portrait_0001.png
├── image_0001.png … image_0050.png
├── flux2_0001.png … flux2_0050.png     (with Flux edits)
└── prompts/                           (after caption export)
    ├── image_0001.txt …
    └── flux2_0001.txt …

The image savers embed the workflow. The base variation saver also enables generation data and previews. Counters may continue from existing files. Connected output paths and filename prefixes control the saved locations.

First-run checklist

Select your installed models and source images; set your own project name and shared specifications.
Mute Flux2 Edit, Image to Prompt and Save Prompts for the first pass.
Keep preparation seeds fixed. Enable face/wardrobe review stops where useful.
Approve the face and wardrobe board.
Choose mannequin Preset settings, or connect physique images for Reference mode.
Enable the character-sheet Stop, approve identity, proportions and clothing, then turn it off.
Check all three auto crops; use manual boundaries for uneven panels. Turn off the portrait Stop after approval.
Generate three test variations in preview mode. Review identity, proportions, clothing, anatomy and shot variety.
Set the intended count and use reserve for the production plan.
Set the optional edit instruction, enable the desired edit/caption branches, and review the saved images and captions.

Files with SHA256: d1f444d16d4fc54f514502627f56eebe9cec53014b4e3a38664c0e5510855a73