Note: This document is intentionally thorough. It may feel long, but it will help you extract the full value from the workflow.
Introduction
This guide documents a fully featured multi-scene AV clip editing workflow that builds on my earlier Poser and multi‑segment scenario workflow collection. While the core ideas remain, there are significant changes. I also neglected to publish usage instructions previously; this document serves as the comprehensive manual.
The workflow now supports quick templated yet customizable access to many scenarios while keeping proven tools at your fingertips for high‑quality, consistent AV videos.
Features
23 scenes in v1.2 (up from 17 in v 1.0)
new - Anal Missionary (3 scenes)
Cowgirl (2 scenes)
Creampie (2 scenes)
new - Deepthroat (7 scenes)
Doggy (4 scenes)
Missionary (4 scenes)
Posing (1 scene).
Some are full stories; others are standalone segments that can be mixed to create long, compelling narratives with multiple poses.
Each scene has a prompt template and recommended LORAs. Override the template or add your own LORAs at will.
Scenes can be generated from either the same source picture or the last frame of the previous scene.
While templates assume one subject, most scenes support multiple subjects, use your imagination.
Disable all templates with a single click and use the same Prompt Override node as your main prompt. The Prompt Override can contain placeholders just like scene template prompts so you can use scene template as a starting point.
Configuration Options - scene selection, video editing, and optimizations are adjustable with one click but fully tunable for any scene.
Torch Compile and Sage Attention - If unavailable, bypass or mute the Torch Compile node and switch to SDPA.
One‑click activation of Painter I2V (improved motion) or default WAN encoders.
One-click activation of LightX2V models: seko v1 for low noise; and a choice of 1022 or 1030 for high noise.
GGUF support - pick the appropriate GGUF model in the loader and disable quantization.
Predefined scene LORAs - add more to the scene LORA stack or a dedicated custom LORA stacks (green nodes)
Portrait/landscape, upscaling, interpolation, and prompt overrides are handled automatically based on your input image.
Keyboard Shortcuts - 1–5 jump to configuration options, inputs, LORAs, etc.
Video Editing Features and Tools
2‑in‑1 I2V & FLF workflow
Improved motion with Painter I2V and high‑LORA options (switch quickly to find the best fit)
Frame cuts - remove unwanted transitions or duplicate frames when stitching a sequence
TensorRT Upscale - 4× upscaling. Far superior to FlashVSR/SeedVR2.
Resize - Upscale standalone using lanczos or downscale after TensorRT Upscale (i.e. from 720p -> 1080p or 720p -> 5K -> 1080p)
RIFE interpolation or fast TensorRT RIFE
Multiple output options: original video, frame image sequence, last frame for next scene, post‑production versions
How to Use the Workflow
The workflow defaults are set so you can start iterating through scenes immediately.
Configuration Options
Tip: Press 1 to quickly navigate to Configuration Options
Processing Options (First Section)
This section contains all post‑processing settings, and you can enable any number of them simultaneously.
First & Last Frame - Switch between a single starting image or a dual first‑and‑last frame workflow. This is handy if you already have two images for a scene, or if you want to reuse the ending frame from a previously generated scene while changing the seed so the story progresses differently but ends as desired.
Frame Skip - Remove unwanted cuts or transitions. For example, in a missionary scenario your subject may transition from sitting/lying to kneeling for a facial cumshot. Skipping those frames turns the transition into a clean scene cut.
TensorRT Upscale - A fast 4× upscale that supports many models (Remacri, ClearReality, UltraSharp, etc.). Results are usually superior to other upscaling methods. The node can also downscale to HD, FHD, 2K, 4K, 2×, or 3×, however, it’s best to leave the default “none” and enable "0 4. Resize" instead. The built‑in node resize uses bicubic, while the workflow’s resize uses lanczos - often a noticeable improvement. You will also have more control over output dimensions.
Resize - A standalone lanczos upscale or downscale that can be applied after TensorRT upscaling or on its own.
Interpolation Options (Second Section)
Only one interpolation method can be active at a time—or none if you prefer no interpolation.
RIFE
TensorRT‑optimized RIFE (super fast)
Note: Current RIFE nodes accept only integer multipliers. I am exploring a new RIFE node that would allow fractional multipliers and possibly GIMM interpolation in the future.
Output Options (Third Section)
Original outputs are generated before frame skip, upscale, and interpolation.
Processed outputs are generated after all post‑processing steps.
For each option you can:
Save every frame as an image.
Save only the last frame (useful as a starting image for the next scene).
Export a video.
If you haven’t enabled any post‑processing, selecting both Original and Processed will produce duplicate files. In that case, simply keep Processed unless you need the unprocessed version, e.g., for additional upscaling in another app or to reduce resolution for efficient LLM/audio generation.
I2V Encoder Selection (Fourth Section)
Only one encoder can be active at a time.
Painter I2V - A newer encoder that improves motion but may introduce composition or anatomical artifacts.
WanVideo I2V - The default encoder.
High‑LORA Version (Fifth Section)
Only one high lightx2v LORA can be active at a time.
10/30 at 0.5 strength - Enhances motion, but like Painter I2V it can affect composition and anatomy.
10/22 at 0.8 strength - Provides solid baseline performance.
Default settings: 10/22 LORA + Painter I2V. This combination generally yields the best results across all scenarios, though certain scenes may benefit from alternative pairings, experiment to find what works best.
Workflow Tips
When you’re first iterating and searching for a good seed:
Keep all processing, interpolation and output options unchecked except Enable Processed Video and optionally Enable Processed Last Frame outputs.
Once you’ve found an acceptable seed, lock it in place. The seed updates after workflow runs so if you have not set it to "fixed" beforehand you will need to copy it from one of output file names.
Enable post‑processing; the workflow will then run only post‑processing steps and outputs, saving you significant time.
Scenarios and Prompting
Tip: Press 2 to quickly navigate to input nodes
Next to the configuration options, you’ll find a list of available scenes. Only one scene can be selected at a time.
Prompt Input Behavior
If a scene is chosen, the prompt override replaces the scene’s template.
If no scene is chosen, the prompt override functions as a plain positive prompt (placeholders still work).
The negative prompt defaults to the WAN list and is adjustable in the Processing section in v1.0. From v1.1 it is right next to positive prompt (override)
Each scene comes with:
Prompt Template - basic templates that can be edited or overridden.
Placeholders - Used to share common information across scenes.
Supported placeholders in v1.1:
Subject:
{subject}, for example "woman"
{subject_description}, for example “beautiful, slender”
{subject_look}, for examples“wavy blonde hair in a ponytail, blue eyes”
{subject_clothes}, for example “wide‑brim hat, white dress” (usually applied to the first frame)
{subject_clothes_adjustment}, for example "She lifts up the hem of the white dress" (conversely, put secret sauce here, expectation is that this placeholder is one or more full sentence(s))
{subject_breasts}, for example “large perky breasts with puffy nipples”
{subject_genitals}, (use responsibly)
{subject_pronoun}, for example “she”
{subject_possessive}, for example “her”
{subject_object_pronoun}, for example “her” (I know subject object sounds weird but by the time I got here I was not ready to rename all placeholders to align with anything else)
Subject State (each state expects one or more full sentence(s))
{subject_inital_state}, for example "She is standing on the beach." (usually applied to the first frame, before any scene cuts, can contain any environment, description, camera control or actions not already covered by look, clothes and other placeholders)
{subject_state_before}, for example "Camera pans out as the woman kneels down." (usually applied before main action starts, and after any scene cut if it exists)
{subject_state_after}, for example "The woman smiles happily" (usually applied after main action ends, although feelings, expressions, etc. may bleed into main action)
Participant:
{participant}, for example "man"
{participant_description}, for example “muscular naked”
{participant_look}, for example “brown hair, green eyes, beard”
{participant_clothes}, for example “nothing” (if description placeholder contains naked)
{participant_clothes_adjustment}, for example "He unbuttons his shirt."
{participant_breasts}, included for consistency but maybe your participant is a woman or just has breasts...
{participant_genitals}, (use responsibly)
{participant_pronoun}, for example “he”
{participant_possessive}, for example “his”
{participant_object_pronoun}, for example “him”
Participant State (each state expects one or more full sentence(s))
{participant_inital_state}, for example "He walks up to her from the right side." (usually applied to the first frame, if participant is already there or to their first appearance., can contain any environment, description, camera control or actions not already covered by look, clothes and other placeholders)
{participant_state_before}, for example "Camera pans out as the man grabs woman by her waist." (usually applied before main action starts)
{participant_state_after}, for example "The man gasps in pleasure" (usually applied after main action ends)
Note: Placeholder format and prompt templates changed significantly in v1.1. If you customized any prompt templates, please update them to use new format.
Most prompt templates follow this general format:
Subject intro (description, initial state, look, clothes, body)
Subject clothes adjustment
Scene switch if exists
Subject before state
Participant intro if exists (description, initial state, look, clothes, body)
Participant before state
Main action
Subject after state
Participant after state
General scene qualities (lighting, etc.)
New placeholder format does not capitalize words, but it has no impact on scene generation. It was more for cosmetic purposes.
Supported placeholders in v1.0:
<subject>, for example “woman”
<subject description>, for example “beautiful, slender”
<subject_pronoun>, for example “she”
<possessive_adjective>, for example “her”
<look>, for examples“wavy blonde hair in a ponytail, blue eyes”
<clothes>, for example “wide‑brim hat, white dress” (usually applied to the first frame; override if you need the subject to stay clothed or partially clothed)
<breasts>, for example “large perky breasts with puffy nipples”
<genitals>, (use responsibly)
<location>, for example “at a beach” or “on a sofa in the living room” (applies to the activity, not necessarily the first frame)
<action>, for example “standing”, “sitting” (again, refers to the activity, not the first frame
Note: (v1.0 ONLY) Capitalized placeholders (e.g. <Subject>) will automatically capitalize the first letter of the value when substituted.
Specific Scene Considerations
Most scenes can be generated from a full body or portrait picture provided you describe camera controls.
Scenes that are numbered (e.g. Missionary and Deepthroat) are supposed to be generated in sequence using last frame of the current step as the first frame for the next step.
If a scene implies a cut and you do not specify a new environment / setting in {subject_state_before}, WAN will often generate a transition instead. It is rarely successful although some results are pretty entertaining / can provide inspiration. So be pretty specific defining new state before action. You can change the setting, for example image starts on a beach but you want action to happen in the bedroom. You can change other things too - clothes, character looks, viewpoint, etc. The sky is the limit.
As you move through numbered scenes do not forget to update the state, looks and clothes, i.e. "before state" or "after state" from step 1 will likely become initial state in step 2 (may need to be adjusted if you described actions or camera controls to just describe the end state of the change). Likewise, if the subject removed or changed any clothes, describe the last frame that now became the first frame for the new step, not the original image.
Deepthroat scenes are easier to start from portrait images. You can crop the full body and optionally upscale the image before using it for the scene. If your intent was to go from full body image, it is ok, but you have to provide camera control (pan / zoom) or make your subject get into appropriate position (kneel, sit, bend over, etc.) otherwise you might get interesting results. Also, if you intended the subject to get naked, you would likely get doll anatomy. The scene template contains LORA for genitals, but the prompt does not use the placeholder for it anywhere. The easiest way to get proper anatomy is to mention genitals specifically in "subject_look" mapping. Otherwise, you can modify the prompt template or use prompt override to get exactly what you need.
If you do not want to deal with prompt templates but ran out of what you can do out of the box, use prompt override.
Most scenes will happily oblige if you want multiple people in it. I recommend using prompt override instead of wrangling with mappings. Make sure you describe each person's look, position etc. or you will get all of them blended together.
In "Missionary - 03. Cum on Body" scene, you can write things in sperm on subject's body.
LORA Settings
Tip: Press 3 to quickly navigate to custom LORA stacks
Each scene defines a set of high and low LORAs.
You can add up to five sets directly to the scene template, plus additional custom LORA stacks (shown in green) that will be used by the workflow between scene switches.
When no scene is selected, you can use the green LORA stacks to specify your own high/low LORAs. Scene‑specific LORAs are toggled automatically and will not interfere.
Key Input Parameters
Short / Long Side (video dimensions)
For best quality:
720 × 1280 -> 16 : 9 or 9 : 16
720 × 960 -> 4 : 3 or 3 : 4
Orientation will be determined by the uploaded image’s aspect ratio. If the image differs from your chosen dimensions, it will be resized and center‑cropped to match.
Dimensions must be divisible by 16, values are automatically rounded down (e.g., 725 -> 720).
Upscale Multiplier
Dimensions are calculated to preserve aspect ratio based on the entered multiplier value.
Duration & Frame Rate
Defaults to 5 s at 16 fps.
Longer videos take more time to generate and may trigger out‑of‑memory errors. WAN models tend to loop back to the original image after 5 s, which can be useful for repetitive actions. You can use 24 fps for smoother motion especially if you’re using the 10/22 LORA with the default WAN encoder.
Skip Frames
If enabled in the configuration options, specify how many frames to trim from the start and/or end of the clip.
Steps
Keep at 8. Change if you feel adventurous and want to experiment with other values.
Shift
8–10 works well.
Lower shift values (e.g., 5) generally produce poor results for this type of video.
Additional Considerations
Negative Prompt
Adjustable in the Processing section (red node).
Schedulers
Adjust blue nodes. In practice, Euler (stable) and UniPC (creative) give the best results.
Adding Placeholders
v1.1
Add new placeholders to one of the four dictionaries and start using them in prompt templates or prompt override.
v1.0
Modify the Placeholder Replace subgraph. It’s a sizable but straightforward stack of “String Capitalize” and “Replace” nodes.
Adding Scenes
Possible, though cumbersome. Each scene is a subgraph that can be copied together with its group (the group must have a specific color to appear in the Scenarios stack). The challenge lies in wiring it into the existing prompt and LORA switches, which are already large although from v1.1 they are at least next to scene definitions.
I actually have a very compact version of the workflow ready to go with Subgraphs but the subgraph functionality is currently so broken in Comfy that I am holding it until it is fixed. Hopefully by the time I am ready to release v1.2.
Future Plans
More Scenes & Better Template Management
I’m exploring ways to manage templates and LORAs more flexibly. I started using Basic_Data_Handling for prompt template and placeholder management. Boyonodes were considered but couldn’t handle LORA strength, an essential feature for some scenes. If you know of better nodes that can manage prompts and LORAs, please let me know.
Additional Tools
Certain options (e.g., auto‑prompting with LLMs, large‑scale video stitching) were omitted deliberately because all‑in‑one workflows often fall short when iteration and clip editing are required.
I plan to integrate artistic filters, color matching, audio generation (mmaudio/Ovi), and other tools in future releases. If you have specific requests, feel free to share them.
Dependencies
Custom Nodes
starting with v1.1
https://github.com/StableLlama/ComfyUI-basic_data_handling
starting with v1.0
https://github.com/kijai/ComfyUI-WanVideoWrapper
https://github.com/rgthree/rgthree-comfy
https://github.com/kijai/ComfyUI-KJNodes
https://github.com/chflame163/ComfyUI_LayerStyle
https://github.com/Kosinkadink/ComfyUI-VideoHelperSuite
https://github.com/Fannovel16/ComfyUI-Frame-Interpolation
https://github.com/yuvraj108c/ComfyUI-Upscaler-Tensorrt
https://github.com/yuvraj108c/ComfyUI-Rife-Tensorrt
https://github.com/princepainter/ComfyUI-PainterI2VforKJ
If you are struggling installation TensorRT nodes, please use this guide
If you are struggling with installation of triton and sage attention please use this guide
https://github.com/Tavris1/ComfyUI-Easy-Install
Models
Text Encoder
https://huggingface.co/eddy1111111/Wan_toolkit/tree/main
WAN 2.2 High and Low models
https://huggingface.co/Kijai/WanVideo_comfy_fp8_scaled/tree/main/I2V
LORAs
LightX2V
10/30
https://huggingface.co/Kijai/WanVideo_comfy/tree/main/LoRAs/Wan22_Lightx2v
10/22
https://huggingface.co/lightx2v/Wan2.2-Distill-Loras/tree/main
Seko v1
Scene LORAs
starting with v1.2
https://civarchive.com/models/1434685
https://civarchive.com/models/2172672
https://civarchive.com/models/2143379
https://civarchive.com/models/2088559
starting with v1.1
https://civarchive.com/models/2078744
https://civarchive.com/models/2109996
starting with v1.0
https://civarchive.com/models/1648982
https://civarchive.com/models/2048863
https://civarchive.com/models/2121111
https://civarchive.com/models/2007166
https://civarchive.com/models/1874811
https://civarchive.com/models/1986500
https://huggingface.co/JustAnotherCibrarian/base_wan22/tree/main/1879839?not-for-all-audiences=true
https://tensor.art/models/902621070467587542
https://civarchive.com/models/1811313?modelVersionId=2176505
https://civarchive.com/models/2048121
https://civarchive.com/models/2131565
https://civarchive.com/models/2118407
Description
What is new in v1.1?
Two new scenes - sitting creampie and standing creampie
New placeholder mapping system, no more struggles with a subgraph, just add more key-value pairs and use them right away
Rewritten prompt templates to support new mapping and also adjust for better flow
UX improvements - related elements, i.e. positive and negative prompts, first and last frame, etc. are now closer together
Scene switches are now closer to scene definitions, slightly simplifying the task of adding new scenes
Known Issues
Power Puter (rgthree) nodes with multiple dynamic outputs, when edited in any way, lose connection with KJ Set Nodes. This causes warnings that Set nodes are missing input that are hard to troubleshoot. In this workflow, the following nodes are affected:
Calculate Dimensions -> Set Height
Calculate Scaling -> Set Height Scaled
The workaround is to delete KJ set nodes, create new set nodes, create a link between Puter nodes and set nodes, collapse puter nodes. Only then name set node variables (height and height_scaled), collapse them and save workflow.