From a prompt to an image comparison

1. Set up Qwen Image

1. Press Tab and add Qwen Image (qwen_image). Choose the backend before opening Setup.
2. Select the model variant and runtime location. For native generation, you can choose a separate parent folder for model storage; Setup creates Qwen-Image-2.1 there and reuses compatible existing checkpoints. Set an output folder if you want a custom save destination.
3. Read the linked model license, choose your permitted model use, and acknowledge it. Choose Install / Verify Selected Model to prepare and verify the selected runtime and weights.
4. Use Start Local Backend, then Check Backend. Generate once the node reports that the selected model is ready.

The native installation runs independently of ComfyUI and does not silently switch to it. Select and set up each backend separately. Download sizes and memory use depend on the selected model, resolution, and memory options.

Qwen Image 2.1 uses the Qwen Research License. The setup acknowledgement does not grant commercial rights; commercial model use requires a separate Qwen license. Model terms are separate from the application and framework licenses.

2. Generate an image

1. Select Text to Image. Enter a description, or connect text to the prompt input. A connected prompt overrides the local field. A single-field Note supplies its value; for a multi-field Note, check the values you want to include.
2. Choose a preset, or set Width, Height, Steps, and CFG. Dimensions use a 32-pixel grid within the supported 64–4096 range.
3. Leave Random seed enabled for variations. Disable it and enter a seed when you want to reuse the same generation settings.
4. Choose Generate. Inspect the preview and status, then use Open PNG or double-click the Qwen preview to open the saved image in your system image viewer. Output Folder opens the save location.

Note data stays available: checkboxes do not remove values from the Note's general output. Qwen Image and all three MiniMax H3 nodes select prompt values at their inputs: multiple visible Note parameters require checks, while exactly one visible parameter is used even unchecked. Only values enter the prompt, without parameter labels. An empty second visible parameter still counts; hidden/internal parameters do not.

3. Edit with input and reference images

1. Select Image Edit. Browse an Input image or connect an image to image. Describe the requested change in the prompt.
2. Optionally supply image 2 through Reference or reference_image. Native generation also accepts additional reference paths, one per line, up to ten images total including the input.
3. For native edits, choose Guidance: Flexible follows the prompt; Edit input favors the original composition; Use input as a pose/layout guide asks Qwen to reuse its composition. These are instruction presets, so exact pose, camera, or clothing preservation is not guaranteed.
4. Choose native Framing independently of Guidance: Automatic uses the last supplied image, Match Input follows image 1, Match first Reference follows image 2, and Custom uses Width and Height. Edit size controls conditioning size and the automatic or matched output pixel budget.
5. Generate and review the saved PNG. Load the reference and generated variants into Image Collection to compare them.

A Qwen image or reference_image connection from Image Collection uses that collection's first path. It does not send the entire collection as multiple native references. Use the reference controls to supply the remaining images.

4. Arrange an image collection

Add Image Collection (image_collection) and choose Load to append images. You can also drag local image files from Explorer into the 2D graph to create a collection. Duplicate paths are removed while preserving image order.

• Preview: single-image sizing follows the source aspect ratio, and multiple images use tightly packed rows. A thin gray outline identifies the image under the cursor.
• Move: click and drag an image to move the collection when Edit is off.
• Resize: drag the node's borders or corners. Images reflow and load sharper previews as the available space grows. The chosen node dimensions are saved with the graph; source files are not resized.
• Edit: enable Edit, select a thumbnail, copy its path or open its folder, and use Remove to take it out of the collection. Removing an entry does not delete the source file. Turn Edit off before opening the viewer or dragging the collection by its images.

The picker offers PNG, JPG/JPEG, BMP, and TIF/TIFF. Graphs store image paths, so keep the referenced files available when moving a project. Compatible consumers include MiniMax H3's multi-image input; Qwen's image inputs use the first collection path as described above.

5. Browse originals in the viewer

With Edit off, double-click a thumbnail to open that image in a large, resizable viewer. It loads the original pixels and initially fits them to the window. The caption shows collection position, filename, and dimensions.

When an image is pinned, every Previous/Next action follows the comparison rules below, including the wheel, side buttons, clicks, swipes, and slideshow.

6. Pin a reference for back-and-forth comparison

The pin button sits immediately left of the navigation triangles. It is gray on an unpinned image and blue when the displayed image is the pinned reference.

1. Open reference A and click the button to pin it.
2. Go Next to comparison image B.
3. Go Previous to see A. Repeated Previous actions stay on A and do not step backward through the collection.
4. Go Next once to return to B. You can repeat Previous/Next to compare A and B as often as needed.
5. From B, go Next again to advance to C. Previous now shows A, and Next returns to C. Forward browsing continues to wrap through the collection.

Pin A → Next: B → Previous: A → Next: B → Next again: C
Previous: A → Next: C

Click the blue pin while viewing A to unpin and restore ordinary backward navigation. Click a gray pin while viewing another image to replace the reference. The pin belongs to the current viewer session.

7. Use designs in a previs-guided video

Add MiniMax H3 Ref Video (minimax_h3_reference_video) from Create Node. It combines a layout/previs clip with character sheets, environment designs, and a prompt through the MiniMax cloud API. MiniMax H3 Local Video uses a local runtime; MiniMax API Vid is the separate text/first-last-frame/image-reference integration.

1. For a rendered image sequence, optionally set [ and ] in Video Player, connect it to Sequence to MP4, and leave Start/End on Auto to inherit those flags. Explicit Start/End numbers override each boundary independently. Both endpoints are included.
2. Supply the encoded clip under Previs videos. Start with one short MP4 using H.264. MP4/MOV and H.264/HEVC are supported: 2–15 seconds per clip, 15 seconds combined, up to three clips, 50 MB each, and 23.976–60 FPS.
3. Connect each character, environment, or other Image Collection to the same Design images pin. Every collection contributes its paths, ordered by graph source position and then each collection's own image order. Duplicate paths are removed. The limit is nine images combined, not nine per collection.
4. Describe each reference's role in the prompt, or connect a Note and check the desired values if it has multiple parameters. For example: “Use reference video 1 for camera movement, framing and poses; reference image 1 for character identity and costume; reference image 2 for the environment.” Reference numbers follow list order within each media type.
5. Enter an eligible MiniMax pay-as-you-go key, or set MINIMAX_API_KEY / MINIMAX_PAYG_API_KEY. Choose H3 (4–15 seconds, 768P/2K) or H3 Max (5–15 seconds, 480P/768P), then Generate. Local inputs upload automatically and generation runs in the background.
6. Connect the resulting mp4_path to Video Player. Its large-player button opens original media in a maximized window with playback controls and a normal title bar. Space toggles play/pause; [X] clears flags.

Optional audio references use WAV/MP3, up to three clips of 2–15 seconds and 15 seconds combined. Use at most 12 files across all media types. See the node reference for dimensions, formats, and file-size limits.

Recover an existing generation: save the workflow to retain its task ID, then use Resume task to query/download without generating again. Stop waiting stops local waiting only; MiniMax continues submitted work and may charge for it. Entered keys stay in app memory and are excluded from saved workflows and task metadata.

Reference following is generative. The model can change camera, composition, timing, poses, or design details. Generated video and reference-video input use provider balance. First/last-frame roles cannot be mixed into this reference workflow.

Troubleshooting

• The backend is not ready: select the intended backend and variant, open Setup, verify the installation and license acknowledgement, then Start and Check Backend.
• Native INT8 or GGUF is unavailable: native currently supports official BF16. INT8 and GGUF Q4_K_M belong to the optional ComfyUI backend.
• Generation runs out of memory: try a smaller preset or Low memory. Native CPU offloading can be much slower; restart ComfyUI after changing its launch profile.
• An edit changes composition or framing: describe each image's role, review Guidance, and choose Framing explicitly. Guidance is not a guarantee of preservation.
• Double-click does not open the collection viewer: turn Edit off. The Qwen node's own preview opens the saved PNG in the system viewer instead.
• Dragging moves the wrong thing: in the graph, dragging an image moves the collection with Edit off. In the viewer, dragging navigates in Fit mode and pans at 100%.
• Previous keeps showing the same image: a comparison pin is active. Return to the pinned image and click its blue pin to restore ordinary Previous navigation.
• A file is missing: locate the original image. The viewer reports the error and lets you navigate to another entry.
• A connected H3 Note prompt is empty: check the desired rows when the Note has multiple parameters. The receiving node will not substitute a saved local prompt for an empty connected selection.
• Too many H3 design images: count all connected Image Collections together. Remove duplicates or reduce the combined list to nine; files are not silently truncated.