Tulut Foot Detailer:
TulutFootDetailer is a ComfyUI node for refining feet in an image. It uses a foot-detection model to locate areas to work on, enlarges each detected area to give the image-generation model more room to add detail, and blends the result back into the original image. The surrounding crop gives the model visual context, while a softened mask concentrates changes near the detected foot and helps avoid a hard edge where the edited area meets the rest of the image. If no eligible area is detected, the image passes through unchanged.
Tulut Foot Detailer Input Parameters:
image
The IMAGE you want to refine. The node detects foot areas in this image, processes eligible crops, and blends the results into a copy of it. There is no numeric range or default; connect an image from an upstream node.
model
The MODEL used to generate the refined detail. The node samples with this model for each eligible detected area. When the optional LLLite application succeeds, it uses a modified model for that area's sampling; it does not replace the model passed through the output. There is no numeric range or default.
clip
The CLIP component used to turn the positive and negative prompt text into guidance for generation. Connect a CLIP compatible with your model. There is no numeric range or default.
vae
The VAE used to convert each enlarged image crop into a form the sampler can process and to turn the sampled result back into an image. Use a VAE appropriate for your model. There is no numeric range or default.
upscale_model
The UPSCALE_MODEL applied to each detected crop before it is resized to the working dimensions. It supplies enlarged image detail for the subsequent refinement step. There is no numeric range or default.
lllite_name
A selectable STRING from the node's controlnet_list. If its value is none, the code skips optional LLLite application; otherwise, it attempts to apply the selected LLLite through an available AnimaLLLiteApply node. The provided source does not list the other menu choices, so the available options depend on the installed setup. There is no numeric range or stated default.
prompt
The positive, multiline STRING describing the detail you want the model to produce in the detected area. It is encoded once and used during sampling of eligible crops. Default: an empty string; no minimum or maximum length is stated.
neg_prompt
The negative, multiline STRING describing qualities you want the model to avoid during refinement. It is encoded once and supplied as negative guidance during sampling. Default: an empty string; no minimum or maximum length is stated.
guide_size
Sets the intended size of the longer side of each working crop before VAE encoding and sampling; both dimensions are rounded down to multiples of eight. A larger value can give the refinement more working pixels but generally requires more processing resources. Default: 512; minimum: 256; maximum: 1024; step: 64.
steps
The number of sampling steps used to generate each refined crop. More steps mean more sampling work and a longer run, but do not guarantee a better result. The sampler and scheduler are fixed in the source to dpmpp_2m_sde and sgm_uniform. Default: 35; minimum: 1; maximum: 100; step: 1.
cfg
Controls how strongly sampling follows the prompt guidance. Higher values put more weight on the supplied text, which may be useful when the desired detail is not appearing but can also change the crop more noticeably. Default: 4.5; minimum: 1.0; maximum: 12.0; step: 0.1.
denoise
Controls how strongly sampling can alter an encoded crop. Lower values generally retain more of its existing appearance; higher values allow more substantial regeneration in the masked area. Default: 0.50; minimum: 0.01; maximum: 1.0; step: 0.01.
character_index
An integer passed to the foot-box detection helper to select a character. The provided context does not include that helper's implementation, so its exact selection order and behavior when an index is unavailable cannot be confirmed. Default: 0; minimum: 0; maximum: 10; step: 1.
seed
The integer supplied to the sampler for each eligible crop. Keeping the seed and other settings the same helps you compare runs under the same sampling conditions; changing it can produce a different result. The same supplied seed is used for each crop in a run. Default: 0; minimum: 0; maximum: 18446744073709551615.
Tulut Foot Detailer Output Parameters:
IMAGE
The resulting IMAGE. Eligible detected areas are refined and blended into a copy of the input image. If no bounding boxes are returned, the node returns the original image unchanged; crops smaller than 16 pixels in either dimension after padding are skipped.
MODEL
The original MODEL input, passed through unchanged so you can connect it to later nodes. It is not the temporary model that may be used when LLLite application succeeds inside the detailer.
CLIP
The original CLIP input, passed through unchanged for use elsewhere in your workflow. Encoding the detailer's prompts does not turn this output into an image or a prompt result.
VAE
The original VAE input, passed through unchanged. The node uses it internally for crop encoding and decoding, then makes the same VAE available to downstream nodes.
UPSCALE_MODEL
The original UPSCALE_MODEL input, passed through unchanged. This output lets you reuse the connected upscaler later without routing another connection from its source.
Tulut Foot Detailer Usage Tips:
- Start with the defaults of
guide_size512,steps35,cfg4.5, anddenoise0.50. Change one setting at a time so you can judge what improves the foot detail. - For a subtle correction that should retain the existing pose and appearance, try a lower
denoisevalue. If the detected area needs more substantial reconstruction, raise it gradually and check whether the result still fits the surrounding image. - If a small detected foot lacks usable detail, try increasing
guide_sizewithin its 256–1024 range. Larger working crops can cost more time and memory. - Write prompts about the detail you want in the foot area. Use
neg_promptfor specific unwanted qualities; both prompts guide every eligible crop in the run. - Leave
lllite_nameatnonewhen you do not intend to use LLLite. Selecting another value makes the node attempt to useAnimaLLLiteApplyif that node is available; an application failure is logged, and processing continues without it. - If the result is unchanged, first check whether a foot is visible enough for detection and whether
character_indexis appropriate. The detection threshold is fixed at 0.50 in this node and is not an exposed setting.
Tulut Foot Detailer Common Errors and Solutions:
Image is unchanged
- Explanation: This is not necessarily an error. When the detection helper returns no foot boxes, the node returns the input image unchanged. It also skips a padded crop if its width or height is under 16 pixels.
- Solution: Check that the foot is clearly visible and large enough in the input image. Try a different
character_indexwhen the image contains multiple characters; the provided source does not establish how that helper orders them.
[Tulut Detailer] LLLite injection skipped: ...
- Explanation: This message is printed when an attempt to apply the selected LLLite fails. The node catches that error and continues sampling with the original model for that crop.
- Solution: Set
lllite_nametononeif you do not need LLLite. Otherwise, verify that the selected option and the installedAnimaLLLiteApplynode work together; consult the traceback printed with the message for the specific failure.
KeyError: 'ImageUpscaleWithModel'
- Explanation: The node looks up
ImageUpscaleWithModelin ComfyUI's node mappings. This error can occur if that mapping is unavailable in the running environment. - Solution: Check that your ComfyUI installation loads its image-upscaling node correctly, then restart ComfyUI and retry the workflow.
Foot detector or model-loading error mentioning FootYolov8x_v20.pt
- Explanation: The node supplies this fixed filename to its bounding-box helper. The helper's implementation is not provided, so its exact loading error and expected file location cannot be confirmed here.
- Solution: Read the full error or traceback, confirm that the required detector model is installed where the extension expects it, and follow that extension's installation instructions.
CUDA out of memory or another memory-allocation error
- Explanation: Upscaling, resizing, VAE processing, and sampling of a crop can require substantial memory, particularly with a large
guide_sizeor large source image. - Solution: Lower
guide_sizetoward 256, use a smaller input image, or free GPU memory used by other workflows before retrying. Reducingstepsmay shorten processing, but is not a substitute for reducing the size of memory-intensive image operations.
VAE, model, or sampler compatibility error
- Explanation: The node passes the connected MODEL, CLIP, and VAE into ComfyUI's encoding, sampling, and decoding nodes. Incompatible components can fail at one of those stages; the exact message depends on the connected components and environment.
- Solution: Use MODEL, CLIP, and VAE components intended to work together. Test them in a basic sampling workflow, then reconnect the working set to this node.
Resize or tensor-size error during crop processing
- Explanation: The node calculates working dimensions from the detected crop and
guide_size, then resizes the crop and mask. An unusual or invalid detected box may produce dimensions that cannot be processed; the detector helper's box-validation behavior is not shown. - Solution: Inspect the full traceback and retry with a clear, sufficiently large foot in the image. If the error persists on ordinary images, report the traceback and example image dimensions to the node maintainer.
