Qwen Pose Direction: Trace the Render, Character Reference, and Active Prompt

Posted on Nov 1 • Originally published at youtube.com

A pose-transfer workflow needs at least two roles to stay clear: the pose guide and the character appearance reference. This export supplies them through different image inputs, while the pose editor also supplies the prompt. That last connection is easy to overlook if you read only the text already displayed inside the encoder.

The supplied package contains one workflow with 14 nodes and 19 links. Its editor state includes three poses and three captured images. Those are not three independent workflow files, nor proof that three final character images were generated successfully.

No model inference, pose-accuracy assessment, identity test, speed measurement or visual acceptance was performed for this guide.

VNCCS_PoseStudio 488 outputs an IMAGE stream at its first port. That output connects to images.image_1 on TextEncodeQwenImage21 484. LoadImage 470 independently connects to images.image_2.

The intended setup therefore separates an editor-produced pose reference from a loaded character image. The third encoder image input is unconnected. Do not describe this as three independent character references simply because the editor stores three captures.

The pose editor itself has no linked inputs in this saved graph. Its pose_data and pose_image sockets are unconnected; its serialized editor state is substantial and may be operationally important. Do not delete that state as though it were only an old comparison thumbnail.

Before adapting the workflow, inspect the actual pose export and use a character image you have permission to process. This guide does not infer the appearance, clothing, identity or content of the selected images from filenames.

The encoder's stored prompt says to replace the pose of image 2 with the pose of image 1 while keeping the character of image 2. That is a useful statement of intended roles, but it is not the active unconnected prompt in this graph.

Pose Studio's second output, named lighting_prompt, connects directly to the encoder prompt input. That link overrides the old widget text.

Draw character from image2 <lighting> <user_prompt> Enter fullscreen mode Exit fullscreen mode Its saved user-prompt field is empty. The three saved lighting-prompt strings are all Draw character from image2. A port called lighting_prompt does not establish that the currently saved strings contain lighting descriptions.

Inspect what the editor actually emits after changes. Do not assume that the stronger pose-replacement sentence inside the encoder is still being sent, and do not conclude from the shorter string alone that pose conditioning is ignored. The first image input and Pose Studio LoRA are also part of the configuration; their actual behavior requires a compatible implementation and a real test.

When adjusting prompts, work with the editor's active output/template rather than editing an overridden encoder field and expecting it to take effect.

The serialized editor state contains three pose records, three captured images and three associated prompt strings. Its export mode is LIST, with saved view dimensions 1520×1024. Reading only the captured PNG headers confirms three stored captures of 1520×1024. This is metadata inspection, not visual review.

The graph then sends the editor image output to GetImageSize 505. Its width and height outputs drive EmptyLatentImage 479, overriding the latent's old 1024×1024 defaults.

However, GetImageSize's batch-size output is not connected. EmptyLatentImage retains batch size 1. Consequently, the capture count alone does not establish a three-image latent batch or exactly three saved result files.

The custom editor's list/batch behavior and ComfyUI execution mapping determine how these outputs are consumed. Verify the installed node behavior and actual file count. Do not advertise an automatic three-pose batch result solely from the saved gallery or the LIST label.

Likewise, saved capture size is not a guarantee that every future editor output or final render remains 1520×1024. Refresh the captures after changing the editor, and inspect effective dimensions and decoded results.

UNETLoader476 → Pose Studio LoRA501 → QwenImage21Cache483 → KSampler481 Enter fullscreen mode Exit fullscreen mode Selected assets are:

The LoRA strength is saved as 1. Cache node 483 stores device gpu and dtype int8. These are configuration references, not independent verification of actual quantization behavior, memory savings, runtime compatibility or a measured speedup.

Resolve the matching custom encoder, Pose Studio and cache-node implementations before queueing. Model filenames and node names are not sufficient installation instructions or license evidence; no download URLs are invented here.

Unlike some other exports, this encoder's VAE socket is connected to loader 478. The same VAE also feeds the final decoder.

Nevertheless, the sampler's initial latent comes from EmptyLatentImage 479, not the encoder's latent output. The encoder's positive and negative conditioning outputs are connected, while its latent output is unused.

The saved KSampler configuration is 25 steps, CFG 1, Euler/simple and denoise 1, with seed 488345170733023 and fixed control. Confirm the controls in the matching UI, including the extra seed-control widget.

This is image-conditioned generation from an empty latent, not a source-latent low-denoise edit or a pixel-preserving deformation of the original character picture. Reference images and a specialized LoRA can express a desired pose/appearance relationship without proving exact joints, identity, costume or anatomy in the result.

There is no separate ControlNet Apply stage or explicit hard joint-constraint input in this export. Assess actual pose correspondence rather than turning the source title's precision language into a guarantee.

Decoded generated image 480 feeds image_a of Image Comparer 504. The editor image output feeds image_b. The loaded character reference at 470 is not the other side of that comparer.

Thus the saved comparison is generation versus pose-editor image, not generation versus the original appearance reference. If identity preservation is an acceptance requirement, inspect the character reference separately or configure an additional reviewed comparison in a working copy.

SaveImage 506 receives the decoded output and uses the prefix ComfyUI. The graph contains one save node; list-driven execution may affect how many files it writes, so do not equate one node with one guaranteed result or three captures with three guaranteed saves.

The historical comparison gallery is cleared in the new canonical copy. Pose Studio's embedded captures and editor state are retained because they may participate in the actual output. Their contents and rights still require review before publication.

First resolve dependencies and confirm that the editor opens correctly. Inspect its three saved captures and replace or refresh them for your intended pose set. Load an authorized character reference, then inspect the actual emitted prompt and image roles.

Check effective width/height, list/batch behavior, model/LoRA/cache settings and save destinations. Test a small case before assuming that every saved pose will map to an aligned prompt and output.

Review pose, hands, silhouette, character identity, costume, lighting and unwanted changes against the appropriate references. A fixed seed is not a general compatibility or reproducibility certificate across node/model versions. No result has been accepted on your behalf here.

The source-provided tutorial supplies context; its link was not live-verified. The source original is preserved. Model weights and the separately selected character-image file are not bundled; the workflow does retain embedded pose captures. Confirm model, LoRA, node, editor-state, character-image and cover rights separately; no universal commercial license or guaranteed pose fidelity is asserted.

Published by this channel. The graph is free; weights, LoRAs and pose assets are separate and carry their own terms.