Ppellucen

Local generation

Qwen-Image-2.1 ComfyUI workflow for transparent images

This guide shows where each part of a Qwen-Image-2.1 ComfyUI graph belongs and which settings are most likely to destroy transparency. Exact model files and node labels can change, so verify them against the checkpoint card and the official ComfyUI template for your installed version.

Before you build the workflow

Use a current ComfyUI build, install the model in the directory expected by its loader, and keep the text encoder and VAE from the same model release. Qwen Image workflows are sensitive to mismatched components because the latent space and decoding path must agree. A workflow that loads without an error can still produce wrong colors or lose the alpha channel.

Start from the official Qwen Image template when possible. Then replace the checkpoint or diffusion model with the Qwen-Image-2.1 files you installed. Avoid downloading a random workflow JSON and pasting model names into it without checking every loader.

Transparency starts in the prompt. Requests that mention a studio backdrop, tabletop, seamless paper, or dramatic floor shadow create an opaque scene even when the graph requests PNG output. Keep the subject isolated and make the background instruction explicit.

For prompt examples and common mistakes, read the Qwen-Image-2.1 prompt guide. If you only need a finished file, use the hosted transparent background generator.

Node-by-node Qwen-Image-2.1 workflow

  1. 01

    Install or update ComfyUI

    Update ComfyUI and its manager before importing a Qwen model workflow. Node names and template availability change, so an older installation may miss the current Qwen Image template or its required frontend features.

  2. 02

    Load the Qwen-Image-2.1 model

    Place the diffusion model files in the folder specified by the model card. Use the Load Diffusion Model or checkpoint node that matches the exact file format. Do not swap in a Qwen 2.0 or 3.0 checkpoint and assume transparent behavior is identical.

  3. 03

    Load the matching text encoder and VAE

    Connect the text encoder, diffusion model, and VAE recommended for the same Qwen-Image-2.1 checkpoint. A mismatched VAE is a common cause of black transparent regions, color shifts, or an export with no usable alpha.

  4. 04

    Write the transparency prompt

    Begin the positive prompt with: This is an RGBA image with transparency. Describe only the subject and its visual qualities. Put background, floor, wall, pedestal, frame, and cast shadow terms in the negative prompt.

  5. 05

    Create a latent or encode an input image

    For text-to-image, use an Empty Latent Image node. For image-to-image or cutout work, load the reference image, scale it, and encode it into a latent with the matching VAE before connecting it to the sampler.

  6. 06

    Sample, decode, and save PNG

    Connect the sampler to VAEDecode, then save with a PNG-capable output node. Do not choose JPEG for the final file. Inspect the saved PNG over a checkerboard and confirm that transparency survived the complete graph.

Reference graph: positive prompt → Qwen text encoder → sampler; empty latent or encoded image → sampler; sampler → VAEDecode; VAEDecode → PNG save. Every connection must use the model’s matching text encoder and VAE.

Troubleshoot transparency and PNG export

If the image looks correct in ComfyUI but the published file is opaque, inspect the final save node first. JPEG cannot store an alpha channel, and some preview or compositing nodes convert RGBA to RGB before the file reaches disk. Save the decoded image directly as PNG and open that file on a checkerboard.

If alpha is present but the subject has a white fringe, the generation may have been influenced by an opaque background. Remove backdrop terms from the positive prompt, strengthen the negative prompt, and test at the resolution recommended by the model card. Downscaling after generation can also make fine edges look harsher.

When local GPU setup is not the goal, compare the graph with the hosted transparent PNG maker or the photo-oriented remove background workflow.

Why is the transparent background black in ComfyUI?

A black or white area often means the workflow is compositing the image or decoding it through a path that does not preserve alpha. Check the VAE, the image preview node, and the final save format. A PNG-capable save node is required for the final asset.

Can I use the same workflow for image-to-image cutouts?

The sampler and decoder can remain similar, but text-to-image uses an empty latent while image editing uses an encoded source image. Start from the official Qwen Image Edit template when one is available for your ComfyUI version.

Do I need the exact prompt from Pellucen?

No, but retaining the RGBA opening sentence and the explicit transparent-background instruction makes the request unambiguous. The rest of the prompt should describe the subject rather than invent a scene.