TakkunandClaude Sonnet 5
ab9d4036cd
- Viewer: send step images as temp files instead of base64 (dead send_json call removed; no history/WebSocket bloat) - Sampler: early stop returns the latest x0 and always saves the stop step; PSNR inf counts as converged - Sampler: ignore auto_stop when stop_threshold is unsuitable for PSNR/SSIM (avoids stopping at step 2) - Sampler: keep step images as uint8 in a local list; pop downscale_ratio_spacial; compat with older fix_empty_latent_channels - Fix display name mapping for the sampler node - Viewer JS: larger slider handle/track, wheel and keyboard stepping, clear timer on node removal, validate play speed - README: fix typos and document threshold semantics / controls Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
72 lines
5.1 KiB
Markdown
72 lines
5.1 KiB
Markdown
# ComfyUI-StepByStep-Sampler
|
|
|
|
A set of custom nodes for ComfyUI designed to analyze and compare the image generation process step-by-step. It quantifies changes in latent variables (using metrics like PSNR/SSIM) and provides a dedicated viewer for intuitive inspection. Furthermore, since sampling automatically stops once image generation converges, image generation can be performed with the optimal number of steps.
|
|
|
|
[](https://github.com/TakkunRed/ComfyUI-StepByStep-Sampler/blob/main/LICENSE)
|
|
[](https://github.com/comfyanonymous/ComfyUI)
|
|
|
|
|
|
|
|
## Node description
|
|
|
|
### 1. Step-by-Step Sampler
|
|
Hooks into the generation process to capture intermediate images at specified intervals by VAE decoding the latent state. It can be integrated with `Preview Image` or `Save Image` nodes to display or output the progression as a sequence.
|
|
Calculates the amount of change from the previous step using MSE, RMSE, L1, PSNR, or SSIM, allowing you to monitor convergence numerically.
|
|
Draws step counts and difference values directly onto the images, making it easy to track the step number and the delta from the preceding frame.
|
|
Since sampling automatically stops once image generation converges, image generation can be performed with the optimal number of steps.
|
|
Outputs the final result as a standard `LATENT` (similar to KSampler), but can also output the VAE-decoded image via the `LAST_IMAGE` socket.
|
|
|
|
<img src="images/StepByStep_Sampler.png">
|
|
|
|
### 2. Step-by-Step Player
|
|
Visualizes the image list output from the `STEP_IMAGES` socket of the `Step-by-Step Sampler`. Use the slider or the Play button to view the generation process as an animation.
|
|
|
|
<img src="images/StepByStepImage.png" width="50%">
|
|
|
|
### 3. Step-by-Step Comparer
|
|
A side-by-side visualization tool for the `STEP_IMAGES` output. It allows you to compare two different steps (e.g., Step 5 vs. Step 20) using an interactive split-screen slider.
|
|
|
|
<img src="images/comparer.png">
|
|
|
|
## Installation
|
|
Manual Install
|
|
```bash
|
|
cd ComfyUI/custom_nodes
|
|
git clone https://github.com/TakkunRed/ComfyUI-StepByStep-Sampler.git
|
|
```
|
|
## How to Use
|
|
### Workflow Integration
|
|
* Add the Step-by-Step Sampler node.
|
|
* Connect `model`, `positive`, `negative`, `latent_image`, and `vae` to the respective inputs.
|
|
* Connect the `STEP_IMAGES` output to the `StepStepPlayer` or `StepStepComparer` node. You can also connect it to `Preview Image` or `Save Image`.
|
|
* Turning on `auto_stop` will stop sampling based on the `stop_threshold`.
|
|
* Run Queue Prompt as usual, and the generation process will appear in the viewers.
|
|
|
|
### Viewer Controls
|
|
* Sliders (Player / Comparer): click anywhere on the track to jump, drag the handle, or hover over a slider and use the **mouse wheel** (down = next step, up = previous step). Arrow keys / `Home` / `End` also work while the slider is focused.
|
|
* `Player`: Change steps using the bottom slider or use the "Play" button for auto-playback.
|
|
* `Comparer`: Select two steps to compare using the sliders, and drag the vertical bar on the image to reveal the differences between side A and side B.
|
|
|
|
### Step-by-Step Sampler Settings
|
|
* `save_interval`: Determines how often images are saved (e.g., set to 1 to capture every step).
|
|
* `show_overlay`: Toggles the on-image display of step numbers and difference metrics.
|
|
* `diff_method`: Selects the algorithm for calculating step-to-step changes:
|
|
* Metric Characteristics (at convergence):
|
|
```
|
|
MSE → 0 (Emphasizes large changes)
|
|
RMSE → 0 (MSE scaled back to L1 range)
|
|
L1 → 0 (Simple, robust against outliers)
|
|
PSNR → ∞ (Higher dB means closer to previous step; 40dB+ is a typical convergence target)
|
|
SSIM → 1.0 (Structural Similarity Index; 0.99+ is a typical convergence target)
|
|
```
|
|
* `auto_stop`: Enable/Disable automatic convergence stop. When it stops early, the image of the stop step is always included in `STEP_IMAGES` (regardless of `save_interval`), and the returned `LATENT` is the state at the stop step.
|
|
* `stop_threshold`: Threshold for determining convergence. The direction depends on `diff_method`:
|
|
* `MSE` / `RMSE` / `L1`: stops when the difference is **at or below** the threshold (e.g. `0.0001`).
|
|
* `PSNR` / `SSIM`: stops when the value is **at or above** the threshold (e.g. `40` dB / `0.99`). If the threshold is clearly unsuitable for the method (PSNR below 10 dB, SSIM below 0.5 — e.g. the MSE-scale default left unchanged), `auto_stop` is ignored and a warning is printed to the console, instead of stopping at step 2.
|
|
|
|
### Notes
|
|
* `STEP_IMAGES` and the overlay use the first image of the batch. `LAST_IMAGE` contains the whole batch.
|
|
* The Player / Comparer viewers receive the step images as temporary JPEG files (`temp/step_by_step/`), not as base64 in the message/history, so long runs do not bloat the server history. Files of the previous run of the same node are removed automatically.
|
|
|
|
## License
|
|
MIT |