一图多视角:
OneImageMultiPrompt lets you explore several edits or interpretations of one reference image in a single workflow. It prepares the image once, then processes each text idea in sequence to create separate guidance for separate sampling branches. This can save memory compared with preparing the same image repeatedly, while making it easier to compare variations from a shared starting setup. It supports several model families, but the way a reference image influences generation depends on the selected model and whether its encoder and sampler support image references.
一图多视角 Input Parameters:
clip
Required CLIP connection. This is the text encoder that turns each prompt, together with the prepared image where supported, into guidance for generation. Connect an encoder compatible with the model family you intend to sample; selecting a model type here does not convert an incompatible encoder.
image
Required IMAGE connection. This is the single reference image reused for every prompt. The node prepares it for visual encoding and, when a VAE is supplied, may also encode it as a latent reference. All active prompts therefore work from the same image, making differences between their results easier to compare.
prompt数量
Required INT controlling how many prompts are processed. Default: 4; minimum: 2; maximum: 8; step: 1. The node processes prompts in order, from prompt1 up to the selected count. Each additional prompt requires another encoding pass, so choose only as many as you need.
vae
Optional VAE connection. When supplied, it encodes the prepared image into latent data. For Qwen Image 2.1 and Krea2, that data is attached to each active conditioning as a reference while sampling starts from an empty latent. For Flux2 and SD / SDXL / Z-Image, the encoded image is also used as the shared sampling latent. Without a VAE, the node omits the encoded latent reference and creates an empty sampling latent; visual-image handling still depends on the connected encoder.
instruction
Optional multiline STRING containing a shared instruction for how the image and text should be interpreted. Its displayed default asks the model to describe key image features, explain the requested alteration, and preserve appropriate consistency with the original. For non-Qwen Image 2.1 modes, it is placed in a system-style template; a blank value uses that same built-in instruction. For Qwen Image 2.1, a custom, non-default instruction is added before each prompt as text, while a blank or unchanged default is not added. This means changing the default matters particularly when you want a shared instruction to affect Qwen Image 2.1 results.
模型类型
Optional model-family selection. Options: Flux2 (default), Qwen Image 2.1, SD / SDXL / Z-Image, and Krea2. This setting changes image preparation, reference handling, output-size calculation, and the shape of the sampling latent. Match it to your actual model, encoder, and VAE; it is not a model loader.
匹配原图尺寸
Optional switch with options OFF (default) and ON. When ON, reference-image preparation uses the source image’s dimensions where applicable, and Qwen Image 2.1 or Krea2 also sizes the output sampling latent to the source dimensions after the mode’s required alignment. When OFF, those paths use the selected output dimensions. For Qwen Image 2.1, this switch does not change the separately prepared visual reference image.
输出宽度
Optional INT specifying output width in pixels. Default: 1024; minimum: 256; maximum: 4096; step: 8. It directly sets the output width for non-Qwen Image 2.1 modes, subject to Krea2’s alignment, unless matching the original image size overrides it. For Qwen Image 2.1, width is normally calculated from 宽高比 and 百万像素; this field is used when the aspect-ratio selection is not one of the defined ratios, such as the manual option.
输出高度
Optional INT specifying output height in pixels. Default: 1024; minimum: 256; maximum: 4096; step: 8. It follows the same rules as 输出宽度 for height. Larger dimensions create larger sampling latents and can substantially increase memory use.
prompt1
Optional multiline STRING; default: empty. Enter the first image-edit or generation idea here. When prompt数量 is at least 2, this prompt is always processed and its guidance appears at cond_1. For example, you could request a warm sunset version of the reference.
prompt2
Optional multiline STRING; default: empty. Enter a second, independent idea for the same image, such as a cool nighttime version for comparison. Because prompt数量 has a minimum of 2, this prompt is always processed and appears at cond_2.
prompt3
Optional multiline STRING; default: empty. This prompt is processed into cond_3 only when prompt数量 is at least 3. Otherwise, its text has no effect on this run.
prompt4
Optional multiline STRING; default: empty. This prompt is processed into cond_4 only when prompt数量 is at least 4, which is the default count. You can use it for another variation without adding a separate image-preparation branch.
prompt5
Optional multiline STRING; default: empty. This prompt is processed into cond_5 only when prompt数量 is at least 5. Increasing the count to include it adds another sequential text-encoding pass.
prompt6
Optional multiline STRING; default: empty. This prompt is processed into cond_6 only when prompt数量 is at least 6. It uses the same prepared reference image as the earlier prompts.
prompt7
Optional multiline STRING; default: empty. This prompt is processed into cond_7 only when prompt数量 is at least 7. If the count is lower, its corresponding output is a placeholder rather than guidance for this text.
prompt8
Optional multiline STRING; default: empty. This prompt is processed into cond_8 only when prompt数量 is 8. It is the last available variation; unused output positions receive placeholders.
宽高比
Optional aspect-ratio selection; default: 1:1 (Square). In Qwen Image 2.1 mode, a defined ratio combines with 百万像素 and 倍数 to calculate output dimensions. If the selection is not a defined ratio, the calculation instead uses 输出宽度 and 输出高度. The supplied source does not list the other ratio labels, so their exact options are not specified here. This setting does not control output dimensions in the other model modes.
百万像素
Optional FLOAT setting the approximate output pixel budget for Qwen Image 2.1 when a defined aspect ratio is selected. Default: 1.0; minimum: 0.1; maximum: 16.0; step: 0.1. The node calculates dimensions from this target and rounds them to the required multiple, so the final pixel count can differ slightly. Higher values generally require more memory; this setting does not size outputs in the other modes.
倍数
Optional INT specifying the dimension-alignment multiple for Qwen Image 2.1 output sizing. Default: 32; minimum: 32; maximum: 128; step: 16. Calculated width and height are rounded to this multiple, with a hard lower limit of 32. This alignment helps produce dimensions suitable for the model’s latent and patch layout; larger multiples make the available sizes coarser.
参考图分辨率
Optional INT controlling the prepared reference-image size in Qwen Image 2.1 mode when 尺寸适配 is 参考图保比例. Default: 1024; minimum: 0; maximum: 4096; step: 32. A positive value targets approximately its square in pixel area while preserving the image’s proportions and aligning dimensions to multiples of 32. A value of 0 keeps the source dimensions as closely as that alignment allows. Values below 256 other than 0 are treated as invalid old-workflow values and reset to 1024. Crop and stretch modes instead size the reference to the output dimensions.
尺寸适配
Optional reference-image fitting mode. Options: 参考图保比例 (default), 参考图裁剪, and 参考图拉伸. For Qwen Image 2.1, preserve-proportions mode sizes the reference independently of the output; crop mode centers it to the output aspect ratio before resizing; stretch mode resizes it directly, potentially distorting shapes. For Krea2, preserve-proportions and crop both center-crop as needed and fit the image to the output grid, while stretch resizes without cropping. This setting is not used by the other model paths.
一图多视角 Output Parameters:
cond_1
CONDITIONING guidance encoded from prompt1 and the shared prepared image. Connect it to a compatible sampler branch for the first variation. If a VAE supplied an encoded reference, that reference is attached to the conditioning; whether it affects generation depends on the model’s reference support.
cond_2
CONDITIONING guidance encoded from prompt2 using the same image preparation as cond_1. Connect it to a separate sampler branch to compare a second idea without preparing the image again. It is active because the minimum prompt count is 2.
cond_3
CONDITIONING guidance for prompt3 when prompt数量 is at least 3. Below that count, this socket contains a zero-filled placeholder, not a meaningful third prompt; leave it disconnected unless you activate it.
cond_4
CONDITIONING guidance for prompt4 when prompt数量 is at least 4. It provides an independent fourth prompt interpretation of the shared image. Below that count, it contains a placeholder.
cond_5
CONDITIONING guidance for prompt5 when prompt数量 is at least 5. Use it for a fifth sampler branch only when that prompt is active; otherwise it is a placeholder.
cond_6
CONDITIONING guidance for prompt6 when prompt数量 is at least 6. It shares the node’s prepared image and sampling latent with the other branches, but carries its own prompt guidance. Otherwise it is a placeholder.
cond_7
CONDITIONING guidance for prompt7 when prompt数量 is at least 7. This lets a seventh sampler branch explore another interpretation; at lower counts, the socket contains a placeholder.
cond_8
CONDITIONING guidance for prompt8 when prompt数量 is 8. This is the final available prompt branch. At any lower count, it contains a zero-filled placeholder rather than encoded guidance.
latent
The shared LATENT sampling input intended for the sampler branches. Its meaning depends on the selected model mode: with a VAE, Flux2 and SD / SDXL / Z-Image use the encoded reference image as the starting latent; without a VAE, they use an empty latent. Qwen Image 2.1 and Krea2 always produce an empty sampling latent, keeping any VAE-encoded reference in the conditioning instead. Separate sampler branches can use this same output as their starting setup.
一图多视角 Usage Tips:
- Set
模型类型to match the model you will sample, then connect a compatibleclipand, when you need latent reference handling,vae. The selection changes processing but does not load or convert model components. - Set
prompt数量to the number of ideas you actually want to run, and connect only the correspondingcond_1throughcond_8outputs. Unused sockets contain placeholders and should not be treated as additional prompt results. - Give each active prompt one clear variation—such as a change in lighting, materials, or setting—while keeping the same reference image. This makes the separate sampler results easier to compare.
- For Qwen Image 2.1, use
宽高比and百万像素as your main size controls. Start with the default 1.0 megapixel budget if memory is limited, and increase it only after the workflow runs reliably. - For Krea2, choose
参考图保比例or参考图裁剪when avoiding stretched subjects matters. Both fit the reference to the output grid;参考图拉伸avoids cropping but can distort the image. - If you want a custom shared instruction in Qwen Image 2.1 mode, change
instructionfrom its displayed default. Leaving it unchanged does not prepend that default text to each prompt in this mode.
一图多视角 Common Errors and Solutions:
Missing required input: clip or image
- Explanation: The node needs both a text encoder and an image before it can prepare and encode prompts. The exact validation wording depends on your ComfyUI setup.
- Solution: Connect a compatible encoder to
clipand a loaded image toimage, then run the workflow again.
Model or encoder shape mismatch during sampling
- Explanation:
模型类型determines latent dimensions and reference preparation. If it does not match your actual model or encoder, a downstream sampler may reject the latent or conditioning with a shape or channel mismatch. - Solution: Select the matching model family and check that the model,
clip, andvaebelong to a compatible workflow. Recreate the sampling connection after correcting the selection.
Out of memory or CUDA out of memory
- Explanation: Large output dimensions, a high Qwen Image 2.1
百万像素value, or large reference-image preparation can exceed available memory. More active prompts also require more encoding work, although the node processes them sequentially. - Solution: Reduce the output size or
百万像素, lower参考图分辨率where it applies, and activate only the prompts you need.
Reference image appears to have no effect
- Explanation: An encoded latent reference is attached only when
vaeis connected, and some model or encoder combinations do not support the reference-image path. The node may log[Toggle-Pass][一图多Prompt]with a warning that the base encoder has no reference capability; it does not change your connections automatically. - Solution: Connect a compatible VAE where required, inspect the console warning, and use a model and encoder workflow known to support the intended reference behavior.
Connected cond_3–cond_8 gives no useful prompt guidance
- Explanation: Outputs above the selected
prompt数量are deliberately filled with placeholders. Text entered into a higher-numbered prompt is not processed until the count includes it. - Solution: Increase
prompt数量to include that prompt, or connect a lower-numbered active conditioning output instead.
Output size differs from 输出宽度 and 输出高度
- Explanation: Qwen Image 2.1 normally calculates dimensions from
宽高比,百万像素, and倍数.匹配原图尺寸set to ON can instead size its sampling latent to the source image; Krea2 also aligns its dimensions to multiples of 8. - Solution: Turn匹配原图尺寸OFF when you want selected output sizing. In Qwen Image 2.1 mode, adjust the ratio and pixel-budget controls, or use the manual aspect-ratio option when you intend to use the width and height fields.
Reference looks cropped or stretched
- Explanation:
参考图裁剪can remove edges to match the output shape, while参考图拉伸can change the proportions of subjects. Krea2’s preserve-proportions mode can also crop to fit the output grid. - Solution: Try
参考图保比例for Qwen Image 2.1, or choose an output aspect ratio closer to the source image for Krea2. Use stretch only when avoiding crop matters more than preserving proportions.
Qwen Image 2.1 custom instruction seems ignored
- Explanation: In this mode, a blank instruction or the unchanged displayed default is not added to prompt text. Only a custom, non-default instruction is prepended.
- Solution: Replace
instructionwith a concise custom instruction, or put the requirement directly into each active prompt.
