ComfyUI_Soze Introduction
ComfyUI_Soze is a collection of custom nodes that helps you manage the practical work around creating images and videos in ComfyUI. Instead of manually opening files, copying prompt text, or repeating the same workflow for every variation, you can use its nodes to load inputs, organize batches, work with text and JSON data, and save results.
It also includes nodes for external services such as ComfyDeploy, FAL, ElevenLabs, and TensorScale. Those services require their own credentials; the everyday file, image, and text helpers serve different purposes and do not all depend on an external service.
The extension originally bundled helpers from the now-abandoned Comfy_KepListStuff project.
Example ComfyUI workflow using ComfyUI_Soze
How ComfyUI_Soze Works
Think of a ComfyUI workflow as a production line: one node supplies an input, another transforms it, and another saves or sends the result. ComfyUI_Soze provides extra pieces for the steps that can become repetitive.
- Feed a workflow from a list. A CSV reader can supply information row by row. If you have a table of prompts, you can use it to work through those prompts rather than entering each one by hand.
- Try combinations systematically. The
CSV Reader X CheckpointandCSV Reader X Loranodes use anindexto walk through combinations of CSV rows and models. For example, a workflow can step through each row paired with each checkpoint. - Prepare and organize media. File and folder loaders bring existing images, videos, and other files into a workflow. Image helpers can then batch, resize, overlay, or otherwise prepare images before the next step.
- Shape information for other nodes. String, JSON, range, and conversion nodes help turn text or structured data into the form another node expects.
- Connect to external services. Integration nodes pass workflow inputs to supported services and bring back their results. These nodes need the relevant service credentials.
You can use only the groups your workflow needs. A simple folder-based image workflow does not need to involve CSV files or an external video service.
ComfyUI_Soze Features
CSV-driven variations
The CSV group includes CSV Reader, CSV Reader X Checkpoint, CSV Reader X Lora, and CSV Writer. These nodes are useful when you want to keep prompts or other workflow values in a table and process them in a repeatable order.
The X Checkpoint and X Lora variants step through rows × models with a single index. This is useful for comparisons: rather than changing both the CSV row and model manually, you can let the workflow advance through each pairing. CSV Writer provides a way to write CSV data from a workflow.
File, folder, and output handling
The Files / folders group includes Load Files From Folder, Load Files With Pattern, File Loader, and Does File Exist. These help you find and load existing material, including files selected from a folder or by a pattern.
For outputs, Save Image With Absolute Filename, Save Text File To Output, and Append To Text File give you ways to save images and text or add text to an existing file. These can be helpful when you want predictable filenames or a running text record alongside your generated work.
Image and video workflows
The Images group includes loaders for individual images, folders, URLs, and file paths: Load Image, Load Images From Folder, Load Images From Folder X Lora, Load Image From URL, and Load Image From Filepath. The X Lora loader is available for workflows that combine folder images with LoRA-based iteration.
Other image nodes include Image Batch Process Switch, Image Overlay, Empty Images, Variable Image Builder, Shrink Image, Pad Mask, Get Most Common Image Colors, Multi Image Batch, and Soze Image Size With Maximum. Together, they offer options for assembling batches, preparing image dimensions and masks, placing images together, and inspecting common colors. For example, size and mask helpers can help prepare inputs before they reach a later image-processing node.
The Video group includes Load Videos From Folder and Append To Video. The latter uses ffmpeg to concatenate video while handling audio; ffmpeg must be available on your system’s PATH for that node to work.
Text, JSON, ranges, and value conversion
The Strings group provides tools such as Multiline Concatenate, Multi Find And Replace, Special Character Replacer, Text Contains, String Functions, String Splitter, Empty String Replacement, Prompt Cache, and Output Filename. You can use these to assemble prompt text, replace repeated phrases, split text into parts, or prepare output names.
The JSON group includes JSON Value Parser, JSON Path Extractor, JSON Array Iterator, JSON File Loader, JSON Formatter, JSON Get Array Count, Create Image Batch From JSON Array, and Load Images From JSONArray. JSON is a structured way to store information; these nodes help a workflow read that information, pick out a value, or work through a list of entries. Image-related JSON nodes can turn image lists in JSON data into workflow inputs.
For controlled variations, the Range / XY group offers Int/Float Range (Step), Int/Float Range (Num Steps), XY Any, and XY Image. These can help when you want to test a sequence of numbers or arrange comparisons across two changing inputs. Converters provide cross-conversion between integer, float, Boolean, and string values, plus Boolean Inverter, so values can be passed to nodes that expect a different type.
Model file loaders and external services
Lora File Loader and Checkpoint File Loader load their respective model files by absolute path. These are useful when a workflow needs to refer directly to a particular model file.
ComfyUI_Soze also includes ComfyDeploy API queue, cache, and download nodes; an ElevenLabs voice retriever; and a FAL Veo3.1 reference-image-to-video node.
The TensorScale nodes run world-model inference synchronously: each node sends one request and streams the finished media back, rather than requiring a separate polling step. Available nodes are:
| Node | What it is for |
|---|---|
TensorScale MiniMax H3 Video (fl2va) | Video with audio. |
TensorScale MiniMax H3 Video Reference (ref2va) | Reference-based video with audio. |
TensorScale LTX-2.5 Fast Video | Video with synchronized audio; the original documentation lists a rate of $0.04 per video second. |
TensorScale LTX-2.3 Fast Video | Fast video generation. |
TensorScale Cosmos3 Nano Text To Video | Video generation from text. |
TensorScale Cosmos3 Nano Video To Video | Video generation using video as input. |
TensorScale Hunyuan Image 3 Fast | Image generation. |
TensorScale SenseNova U1.5 | TensorScale inference using SenseNova U1.5. |
For reference media, TensorScale nodes include the media in the request by default. Because the complete request body is limited to 10 MiB, video-sized inputs may be too large. In that case, use the relevant *_url input or enable use_fal_upload. See the TensorScale API documentation (https://www.tensorscale.io/docs.html) for service details.
ComfyUI_Soze Models
ComfyUI_Soze is primarily a node collection, not a downloadable model pack. Its checkpoint and LoRA loaders help you use model files you already have; the extension does not specify a required checkpoint or LoRA.
The named MiniMax H3, LTX, Cosmos3 Nano, Hunyuan Image 3, and SenseNova U1.5 options above are TensorScale service nodes. Choose one according to the kind of result you need: for example, a text-to-video node for a video prompted from text, a video-to-video node when you have source footage, or an image node for still images. FAL Veo3.1 provides a separate reference-image-to-video integration.
Troubleshooting ComfyUI_Soze
An external-service node reports a missing key or authentication error
-
Identify the service used by the node.
-
Check that its credential has been provided under the expected name:
Service or feature Credential ComfyDeploy CD_API_KEYElevenLabs voice retriever ELEVENLABS_API_KEYComfyDeploy image parameter upload SOZE_AZURE_STORAGE_CONNECTION_STRINGFAL FAL_KEY, or[API] FAL_KEYinconfig.iniTensorScale TENSORSCALE_API_KEY, or a model-specific key -
Credentials can be supplied as environment variables or in a
.envfile in the repository root, as described by the original documentation. After correcting a credential, retry the node. -
Keep keys private. The repository’s
.gitignoreexcludes.envandconfig.ini, but you should still never commit credentials. Rotate a key if you suspect it has leaked.
For TensorScale, model-specific keys fall back to TENSORSCALE_API_KEY. The supported suffixes are MINIMAX_H3, LTX2_5_FAST, LTX_2_3_FAST, COSMOS3_NANO, HUNYUAN_IMAGE_3, and SENSENOVA_U1_5; for example, TENSORSCALE_API_KEY_MINIMAX_H3. TENSORSCALE_BASE_URL can override the service host.
Append To Video cannot run
- Confirm that
ffmpegis installed. - Confirm that the command
ffmpegis available on your system’sPATH. - Retry the workflow once ComfyUI can access it. This node calls
ffmpegto concatenate video and handle audio.
A TensorScale reference-media request is too large
- Check whether you are sending a large image or video as reference media.
- Remember that the entire request body has a 10 MiB limit when media is included directly.
- Use the node’s appropriate
*_urlinput or turn onuse_fal_uploadfor video-sized media, then try again.
Frequently asked questions
Do I need an external-service key for every ComfyUI_Soze node?
No. Keys are needed for the corresponding service integrations, not simply because you use a local file, image, CSV, or text helper.
Does ComfyUI_Soze include checkpoints or LoRAs?
The documentation describes loaders for checkpoint and LoRA files by absolute path, but does not list bundled model files.
How can I test every CSV row with multiple models?
Use CSV Reader X Checkpoint or CSV Reader X Lora, depending on which type of model you are comparing. Their index moves through the combinations of rows and models.
Why should I use a URL for TensorScale reference video?
Reference media is included directly in the request by default. A video can push the request over the 10 MiB limit, so a *_url input or use_fal_upload is more suitable for larger media.
Learn More about ComfyUI_Soze
- ComfyUI_Soze repository — browse the project and its available nodes.
- TensorScale API documentation (https://www.tensorscale.io/docs.html) — details for workflows using the TensorScale service nodes.
- Comfy_KepListStuff — the earlier helper project whose nodes were originally bundled into ComfyUI_Soze.
