This node pack has two sets of nodes:

• Render Pass Nodes
• WAN 2.2 Nodes

Render Pass Nodes

These nodes aim to enable workflows involving complex prompt randomization, and chaining the renders of multiple different checkpoints together.

You can drag this image into ComfyUI for an example workflow showing off most of the features.

Intended Workflow:

• Chain together Render Pass nodes with different checkpoints. Eg., one model might work well with a LoRA, but you want to use the style of another model. Or one anime checkpoint is great at generating a pose you like, but you want to then render over it with a realistic checkpoint.
• Randomize your prompt in a highly controlled way by defining "tags", which you can reference in your prompt to replace a keyword with all the prompt words that describe it. For example, once you define a prompt which describes a character, you can always bring that character into your prompt with a single keyword, eg. <miku>. You can also define random variations within these tags, such as different outfits, hairstyles, etc, with the usual {option a|option b} syntax.

Render Pass

This node includes a sampler, and is designed to be chainable, to allow conveniently building a workflow which chains multiple checkpoints together. For example, your first Render Pass could take an Empty Image as an input, and fully replace it with noise (txt2img), and then a 2nd render pass can take the result, upscale by 50%, and add 50% noise (img2img).

Face Render Pass

Same as Render Pass but uses the FaceDetailer implementation from Impact Pack/Subpack. You need both of those installed.

Prepare Render Pass

Combines data from the below nodes, and lets you provide a CivitAI API key (which you can get here) which allows automatic download of missing LoRAs. This does not work with Checkpoints due to frontend limitations. It also provides two seed values; One for the noise, and one for the prompt randomization. All of this combined can provide the Render Pass node with everything it needs to make your prompting and workflow sharing more convenient.

Built-in tags:

• <!>this text</!> will be moved to the top of the prompt, making it more likely that it will be adhered to.
• <neg>this text</neg> will be moved to the negative prompt.
• <face>this text</face> will be sent to the Face Render Pass node's prompt.
• You can also define arbitrary tags, for example, let's say you want to be able to exclude eyes without having to delete them from the prompt every time: <eye>blue eyes</eye>. Then, you can include <!eye> in your prompt, which will cause the contents of any tags to be excluded from the final prompt.
• <ratio:landscape>/<ratio:square>/<ratio:portrait> will override your image resolution by flipping or averaging your starting image's width/height.
• <!modelprompt> will exclude the base prompt associated with the checkpoint, which is defined in the Prepare Checkpoint node, and is normally prepended to your prompt.

Tag Stacker

Make a single "tag" available in your prompt. For example, lets you define a tag called random clothes with the contents {jacket|shirt|tanktop|bikini}. Later, if you type <random clothes> in the prompt for the RenderPass node, it will replace that tag with the provided contents.

• This node can be chained together to build up a database of tags.
• Tags can also include tags (Avoiding infinite loops is up to you!)

Tag Tweaker

Search & replace in your tag database. Useful if you find yourself typing out combinations of prompt elements often, but don't want to define a tag for them, because you don't want to use them in your final prompt. Eg., you can replace blink with blink <!eye>, which in combination with the <eye> tag example mentioned above, can let you add blinks without having to manually remove the eye-related words from your prompt.

Prepare Checkpoint

Can be chained together just like Tag Stacker to define a database of checkpoints. Inputs:

• (required) CivitAI model ID
• Model version name
• Override filename
• Positive/Negative prompts associated with this checkpoint. This allows you to switch to a different checkpoint without having to change your quality tags, eg., so you can switch between Pony and non-Pony models more easily.
• CFG, Steps, Scheduler, Sampler, the usual suspects.

These values will be used by the RenderPass node when the corresponding checkpoint is selected.

Prepare LoRA

Same as Prepare Checkpoint, but for LoRAs. You can activate LoRAs by adding them to your prompt with this syntax: <lora:MyLora:1.0>. You do not need the full filepath, nor the file extension, although only .safetensors is supported. The file just needs to exist in your ComfyUI/models/loras/ directory. If a LoRA is not present but its model ID is provided through this node (and an API key is provided), it will be automatically downloaded.

Adjust Image

Simple Brightness/Contrast/Saturation adjuster. I highly recommend reducing all 3 between render steps, especially for night-time images.

Misc Nodes

The functionality of these nodes is built into the RenderPass node, but you can also use them outside of that framework, if you so wish.

• CivitAI Model Downloader: I wasn't happy with the one that's online so I made my own. You can enter a model ID, your API key, a desired folder path, file name, and even model version name, and it will download it. Will automatically place checkpoints and LoRAs in the appropriate folders.
• Stable Random Choice: I wasn't happy with how ComfyUI or WildCardProcessor handles randomized elements of a prompt, such as {red|green|blue}, particularly when a lot of such randomized elements were in a prompt. When you modify the prompt without adding or removing any randomized elements, you might get a totally different result. Not with this node. Only if you add or remove random elements.
• Tidy Prompt: Adds commas at end of lines, removes lines starting with #, remove excess commas.
• Extract Tag From String: Kinda like an HTML tag, eg. if the input string is a girl with <eye>blue eyes</eye>, this node will extract the blue eyes part if you ask it to extract the eye tag. Useful for not having to jump between so many different input boxes for prompts, since you can eg. use to send stuff to your negative prompt or to send stuff to your FaceDetailer prompt.
• Auto Extract Tags From String: Good to run at the end of your other tag processing nodes, to remove any leftovers. Additionally, if the input contains a tag with an ! like <!eye> then the contents of any tags will be removed. You could use this to replace blink in your prompt with blink <!eye>, so when you want to quickly make your character blink, you can just add "blink" to your prompt and it will automatically remove any descriptions of eyes. (As long as your eyes are tagged properly)
• Regex Operations: Exactly that.

WAN 2.2 Nodes

You can drag this image into ComfyUI for an example workflow showing off most of the features.

Tired of juggling two models, two LoRA stacks, and two samplers just to run a single WAN 2.2 generation? These nodes help de-duplicate these noodles, and compact a complete WAN 2.2 video generation workflow with LoRAs into an elegant small node set-up.

WAN 2.2 Dual Model

Create a complete model definition for WAN 2.2 by choosing the high and low noise models. You also don't need to worry about whether it's a GGUF model or not. However, if you do choose a GGUF model, you do need to have ComfyUI-GGUF installed, otherwise an error will tell you so.

WAN 2.2 Dual LoRA

Bundles the high and low noise halves of LoRAs together with an activation keyword into a single output. The keyword is used to reference this pair in prompts via one of these syntaxes:

• <lora:keyword> - The LoRA with the matching keyword will be applied to both high and low noise models with a weight of 1.0.
• <lora:keyword:weight> - The LoRA with the matching keyword will be applied to both high and low noise models with the specified weight value.
• <lora:keyword:high_weight:low_weight> - The LoRA with the matching keyword will be applied to both high and low noise models with the corresponding weight value.

WAN 2.2 LoRA Stacker

Accepts a dynamic number of Dual LoRA inputs — a new input slot appears automatically each time you connect one. Outputs a list of all connected LoRA pairs for use by the Render node.

WAN 2.2 Render

The main rendering node. Handles the full image-to-video pipeline internally:

• Rescales the start image to the chosen resolution while preserving aspect ratio
• Loads the CLIP encoder, VAE, and both high and low noise models
• Parses and applies any LoRAs referenced in the prompt
• Encodes prompts and prepares latents
• Runs the high-noise sampling pass, then hands off to the low-noise pass at the step defined by the switch factor
• VAE-decodes and returns the final frames

Outputs decoded video frames and the FPS value for use by downstream nodes (e.g. a video combine node, or a frame interpolation node).

SAGE Attention: If you have ComfyUI-KJNodes AND SageAttention + Triton installed, the Render node will automatically attempt to patch the models with SAGE Attention. This can result in a significant speed boost, in the case of my RTX 5080 about a 33% faster generation for identical results.