Eric_Qwen_Edit_Experiments Introduction
Eric_Qwen_Edit_Experiments is a collection of custom nodes for ComfyUI. It gives you tools for two related kinds of image-making:
- Editing an existing image with Qwen-Image-Edit. You can describe a change, combine references, transfer a style, or work on a masked area.
- Creating an image from a prompt with Qwen-Image. You can choose a canvas shape, build an image in stages, and optionally guide its structure with a ControlNet image.
The collection also includes tools for LoRAs, prompt rewriting, faster sampling, masks, and high-resolution output. You can start with a simple loader and one edit or generation node, then add controls as your project calls for them.
Example high-resolution image edit
The author provides an advanced Qwen-Edit workflow preview. The original README also points to a basic workflow in the repository’s workflows folder.
How Eric_Qwen_Edit_Experiments Works
Think of a pipeline as your working set of image-making tools: it contains the model and the parts needed to interpret your prompt and produce an image. A loader node prepares that pipeline; an edit or generation node uses it.
The two workflows have different starting points:
| What you want to do | Start with | Then use |
|---|---|---|
| Change a photo or illustration | A Qwen-Image-Edit loader and an input image | An edit, inpaint, fusion, or style-transfer node |
| Create an image from words | A Qwen-Image loader | A Generate, Multi-Stage Generate, or UltraGen node |
The editing and generation pipelines are separate. An edit node expects a QWEN_EDIT_PIPELINE; a generation node expects a QWEN_IMAGE_PIPELINE. Their connections are not interchangeable.
When editing, the model can use two kinds of information from an input image. The visual-language path helps it understand what is in the picture, while the image-reference path supplies more direct visual detail. Some multi-image and style nodes let you choose which path receives each image. For example, you can show the model a style reference so it understands the look you want without asking it to reproduce that reference’s exact pixels.
The nodes run the Hugging Face diffusers pipeline with its model-specific scheduling. You do not need to add a separate “Aura Flow Shift” node. Edit nodes aim to retain the input image’s resolution, aligned to the model’s requirements, unless you set a lower max_mp limit. That limit is a safety cap in megapixels: setting it lower reduces the size of a large output and can help with memory use.
For high-resolution creation, multi-stage nodes work much like a painter making a composition sketch before adding detail. Stage 1 establishes the scene at a smaller size; later stages enlarge and refine it. This gives you separate control over the broad composition and the finishing passes.
Eric_Qwen_Edit_Experiments Features
Editing and combining images
- Eric Qwen-Edit Image makes a prompted change to one image. The node defaults to 8 steps, which the author associates with a lightning LoRA; the original README suggests 50 steps for the base model.
max_mpcaps output size, whileseedlets you repeat a result when the other settings stay the same. - Eric Qwen-Edit Inpaint works on a mask: white marks the area to change, and black marks the area to keep. It blanks the selected area for the model, then blends the generated result back into the original.
mask_modechanges how the blank area is presented, andfeathersoftens the join. This is a masked-editing approach, not a native mask input to the edit model. - Eric Qwen-Edit Inpaint Transfer places content from a second image into a masked area, then asks the model to harmonize it with the surroundings. An optional
transfer_maskselects only part of the source.blend_strengthcontrols how strongly the source is placed before harmonization;feathercontrols the final edge blend. - Eric Qwen-Edit Multi-Image Fusion combines two to four images. Choose a
composition_modesuch asgroup,scene, ormerge, and select themain_imagethat anchors the output size. Per-imagevlandrefswitches let you treat one image mainly as an idea or subject reference and another as a stronger pixel reference. - Eric Qwen-Edit Style Transfer applies the look of a
style_imageto acontent_image. Modes include full style, color palette, lighting, artistic medium, and texture. You can add guidance or supply acustom_prompt, which overrides the mode’s prompt template. The style image’s pixel-reference path is off by default to help avoid unwanted details carrying over.
Generating images
- Eric Qwen-Image Generate creates an image from a prompt. It offers nine resolution presets, including square, portrait, and landscape shapes, plus custom width and height.
max_mpcaps the output size; a cap below your chosen preset may reduce the result. - Eric Qwen-Image Multi-Stage Generate offers up to three stages, each with its own step count, guidance strength (
CFG), and refinement settings. Setupscale_to_stage2to0to stop after Stage 1, orupscale_to_stage3to0to stop after Stage 2. - Eric Qwen-Image UltraGen is a quality-focused multi-stage option with tuned defaults, per-stage seed choices, selectable refinement schedules, longer prompt support, and optional VAE upscaling. Its
max_sequence_lengthcan be set from 128 to 1024 tokens for detailed prompts. Itsseed_modecan reuse a seed, offset it between stages, or choose a new random seed for each stage. - Eric Qwen-Image UltraGen CN adds ControlNet guidance to multi-stage generation. Supply a suitable edge, depth, or pose control image to influence the arrangement of the result while your prompt describes its appearance.
cn_target_strengthadjusts the guidance when automatic scaling is enabled; Stage 2 guidance can be reduced or disabled separately.
CFG controls how strongly generation follows the prompt. The nodes use two model passes per step for this guidance, so stronger guidance can also increase processing work. Denoise strength controls how much a later stage is allowed to remake its input: lowering it generally favors keeping the earlier composition, while raising it allows more change. These are creative controls, not guarantees of a particular result.
Detail, speed, and creative control
- LoRA nodes apply compatible LoRA adapters to edit or generation pipelines. LoRAs can introduce a trained style or concept; the available effect depends on the adapter you choose. You can chain adapters with individual weights and use Unload LoRA to return to the base pipeline. A LoRA intended for one pipeline should not be assumed compatible with the other.
- Spectrum Accelerator nodes offer training-free sampling acceleration for editing and generation. They predict some intermediate work instead of running the full transformer on every step. The author reports approximately 3–5× speedups, with possible quality tradeoffs. Spectrum automatically disables itself below its
min_stepssetting, so an 8-step edit will not benefit from the default 15-step threshold. - Eric Qwen Prompt Rewriter expands a short idea into a more descriptive prompt through a local or remote OpenAI-compatible language-model service. You can choose English or Chinese, adjust creativity with
temperature, and add instructions.passthroughlets you compare the original prompt with a rewritten one. If a LoRA needs trigger words, the node can incorporate, prepend, or append them. API keys are read from environment variables orapi_keys.ini, rather than stored in the workflow JSON. - Upscale VAE support can enlarge UltraGen output during decoding using the optional Wan2.1-VAE-upscale2x model.
final_decodeproduces a 2× final decode;inter_stageuses the upscale between stages;bothcombines them for a stated 4× total upscale. Output size and memory needs can rise substantially.
Finishing and memory tools
- Eric Qwen-Edit Delta Overlay detects differences between an original and an edited image, then applies the changed areas back onto the original. Raise
thresholdto ignore smaller differences; adjustblur_radiusandexpand_maskto soften or widen the applied region. It also outputs the detected change mask for inspection. - Eric Qwen-Edit Apply Mask blends a foreground over a background: white mask areas show the foreground, and black areas show the background.
blur_masksoftens the transition. - Unload nodes free an edit or generation pipeline when you are finished with it. Loader options such as VAE offloading, attention slicing, and sequential offloading can reduce memory pressure, generally at a speed cost. The original README gives approximate VRAM guidance of 24 GB for up to 2 MP, 48 GB for up to 6 MP, and 96 GB for up to 16 MP; treat these as guidance rather than a promise for every workflow.
The Component Loader nodes are advanced options for trying compatible fine-tuned or quantized Qwen components. For most artwork, use the standard loader. Component loaders still require a base pipeline for shared configuration, and unrelated Stable Diffusion or CLIP components cannot be substituted.
Eric_Qwen_Edit_Experiments Models
| Model | Best use | Notes |
|---|---|---|
| Qwen-Image-Edit-2511 | Prompted edits, fusion, style transfer, and the edit nodes’ masked workflows | Recommended edit model in the original README. It also mentions Qwen-Image-Edit-2509 as an alternative. |
| Qwen-Image-2512 | New images from text, including multi-stage generation | Recommended generation model in the original README. |
| Qwen-Image | New images from text | Listed as another supported generation choice. |
| InstantX Qwen-Image ControlNet Union | Guided generation from Canny edges, soft edges, depth, or pose | Optional model for UltraGen CN. Choose the cn_type that matches your control image. |
| InstantX Qwen-Image ControlNet Inpainting | Experimenting with mask-guided generation | Optional experimental model. The author reports visible halos and ghosting at mask boundaries in the inpainting workflow. |
| Wan2.1-VAE-upscale2x | Larger UltraGen output during VAE decoding | Optional upscale decoder, not a replacement for the main edit or generation model. |
A LoRA is an additional adapter applied to a compatible main model, not a standalone replacement for it. The effect and appropriate settings depend on the particular LoRA.
Troubleshooting Eric_Qwen_Edit_Experiments
A node will not connect to the loader
- Check whether the node name says Qwen-Edit or Qwen-Image.
- Connect Qwen-Edit nodes to an edit loader and Qwen-Image generation nodes to a generation loader.
- For a LoRA, use the LoRA node matching that same pipeline type.
The two pipeline outputs have different types, even though both use Qwen models.
A large image runs out of memory
- Lower
max_mpon the edit or generation node, or choose a smaller resolution preset. - For a multi-stage workflow, reduce the size or number of stages.
- Try loader options such as
offload_vae,attention_slicing, orsequential_offload, allowing for slower runs. - Unload a pipeline after its last use. Large upscale-VAE outputs may require substantially more memory.
Spectrum does not seem to speed up a run
- Compare runs with the same prompt, resolution, steps, and other settings.
- Check
min_steps: Spectrum is disabled below that count. - Try a workflow with more sampling steps. The generation Spectrum node is described as most useful at 20 steps or more.
A masked edit has a visible border
- Confirm that the mask selects the intended area: white changes; black stays.
- Adjust
featheron an inpaint or transfer node to soften the edge. - For an inpaint transfer, check that the source content fits the target region and adjust
blend_strength. - Keep expectations measured: the Qwen-Edit inpaint node processes the image internally before blending, so it is not the same as generating exclusively inside a mask.
ControlNet inpainting produces halos or ghosting
This is a known limitation in the experimental ControlNet inpainting workflow described by the author. If you need a masked edit now, try Eric Qwen-Edit Inpaint instead. The author reports better results from that approach, while noting that it too is experimental.
A ControlNet-guided result does not follow the intended structure
Check that cn_type matches the control image you supplied: for example, select depth for a depth map rather than a Canny edge map. In UltraGen CN, you can then adjust cn_target_strength with auto-scaling enabled, or inspect whether the guidance start, end, and Stage 2 settings limit when the control image is used.
Frequently asked questions
Do I need a source image to make new artwork?
No. Use a Qwen-Image generation node and a text prompt. Source images are required for the relevant Qwen-Edit workflows.
Will every edit keep my original image size?
The edit nodes are designed to preserve input resolution, subject to model alignment and your max_mp cap. A large input above that cap will be reduced.
What is the simplest way to begin?
For editing, connect an edit loader, an input image, and Eric Qwen-Edit Image. For a new image, connect a generation loader and Eric Qwen-Image Generate. Add multi-stage or reference controls once you have a basic result.
Can I use my usual ComfyUI model-loading nodes?
Not for these pipelines. The original README states that standard ComfyUI “Load Diffusion Model,” “Load CLIP,” and “Load VAE” outputs are not compatible with these nodes’ diffusers pipeline inputs.
Learn More about Eric_Qwen_Edit_Experiments
- Browse the repository’s example edit image and advanced workflow preview to see the kinds of results and node arrangements the author describes.
- Use the model pages linked in the Models section to learn which source model or optional guidance model belongs to a workflow.
- For general help understanding the canvas and node connections, see the ComfyUI documentation.
