A simple workflow for Klein family models in the first row was developed for 9b and 4b distilled versions, but it should work for base models with minimal changes (such as replacing ConditioningZeroOut with negative conditioning).
IMPORTANT!
DemonAlone-JS_Addon-ComfyUI is stuck on version 1.5.3, where I haven't yet added my toggles, and the actual version is v1.9.0. You need to install the nightly version via the ComfyUI manager if your security settings allow it (or change them), or you can go directly into the extension folder and run
git pull origin main from the terminal.
I am tired of beating the dead horse regarding Legacy mode and have decided to try removing Nodes 2.0. However, I didn't anticipate how deep that issue runs. Anyway, here is a version for you.
In theory, everything (except comparison previews) should work in Legacy mode too. But there is a kind of issue: my status toggles will switch nodes into Full mode (too many nodes), if you switch the interface from Legacy to Nodes 2.0 with an open workflow. You just need to toggle it via the context menu between Full mode and Compact mode.
Since for three months, despite my request, I haven't received a single feedback on the 🚩vX Latent versions of this workflow, I've concluded that it's unnecessary or pointless. 🚩v2.5 Latent will remain the last public experimental release and stay private. Ordinary versions will continue to be released as usual.
🚩vX Latent This is an experimental versions of the workflow. It’s not intended for general use (though it might still work as a regular one🟢). The main goal here is to slow down degradation when editing images individually by one-by-one editioning. Typically, using VAE decode artifacts appears after 3–5 edits, which can be addressed by simply removing the latent encoding/decoding steps via two nodes: Latent Loader and Latent Save. This approach fully bypasses VAE entirely for preview purposes—only to use it during editing previews.
I implemented a chain of K-samplers and used a specific schema; while there are no visible artifacts, over time the colors fade slightly, but this is still better than the standard Edit mode. If you try it out, please provide feedback about this workflow (whether positive or negative). I recognize that other methods exist for restoring images, but they’re much more complicated—so I’m exploring this approach first.
At least for now, changing the size in Latent Edit mode isn’t available. It’s possible, but it comes with many limitations; I didn’t find the way I like. So you’ll have the same size as what was saved (.latent file).
Description
This is a fork or spin-off from a regular version, featuring latent load/save functionality.
FAQ
Comments (8)
Hi! 😊 Thank you so much for this workflow! I tested the IMG2IMG mode so far, and it works really well. The result is surprisingly good, especially with character LoRAs - it changes the face/identity while still preserving the original image structure, which is exactly what I needed.
I also really appreciate the overall design and the extra features. The workflow is powerful and very useful once you understand how it is organized.
I wanted to ask a couple of questions, though.
First, about the Resolution Master node: on my side, its UI seems to be glitching. The node appears almost like a blank green block, and the parameters are not editable from the side panel. I can only accidentally click invisible areas inside the node, which makes it very easy to change something unintentionally, for example batch size. Would it be possible to expose the main parameters like width, height, and batch size in a more accessible way, or make them easier to edit from the side panel?
Second, I wanted to ask about the negative prompt. I could not find a clear negative prompt field in the workflow. Did you design it to work without a negative prompt by default, or is there a specific place where it should be connected? If I add my own negative prompt manually, will it actually affect IMG2IMG/Edit/Inpaint modes correctly?
Other than that, thank you again for sharing it. The workflow is really cool and already helped me get much better results than my previous attempts👍🏻
@3d_irin, Thank you for such nice feedback!
About Resolution Master: You are probably using Nodes 2.0, and I experience this issue as well. I don't use Nodes 2.0 or intend to do so in the future. Nonetheless, resolution can still be changed on core nodes (for latent size and width/height).
Regarding negative conditioning, yes—I don’t use it since I’m not using Base versions. However, if you want to enable it under UNet Loader with the "Name" (Image Saver) node—there’s a green subgraph called “Positive” (which I now realize is named incorrectly), you’ll need to remove the red ConditioningZeroOut node inside that graph and replace it with a similar copy of CLIP Text Encode (Prompt). Connect clip input to out, then connect negative conditioning to its output.
If you run into any issues with either the first or second problem, feel free to text me back; I’ll do the necessary workflow modifications for you.
Nice workflow but as of a few days ago getting this error message; Return type mismatch between linked nodes scheduler, received_type(['simple', 'sgm_uniform', 'karras', 'exponential', 'ddim_uniform', 'beta', 'normal', 'linear_quadratic', 'kl_optimal', 'bong_tangent']) mismatch input_type(['simple', 'sgm_uniform', 'karras', 'exponential', 'ddim_uniform', 'beta', 'normal', 'linear_quadratic', 'kl_optimal', 'bong_tangent', 'beta57'])
@three7s Usually, this involves either an installed RES4LYF or another custom sampler/scheduler. If you still want to use them and avoid the error message, go inside the KSampler subgraph and delete both instances of the 'Any to String (Image Saver)' node. Your images won’t have metadata about the sampler/scheduler anymore, but the error should disappear.
Nice! Works fine with RES4LYF disabled. Thank you.
Thanks for the best workflows. You have the same workflow in archives v3.3 Latent and v3.3 Manual inpaint.
Could you please explain in more detail how to use Inpaint and Latent Edit in this workflow? I can't quite figure it out. When using Latent Edit I get an error. And for Inpaint how do I draw a mask?
And regarding controlnet. What query should I use in the hint to activate it correctly?
@volianticq5953 , Thank you for the information about the file. The JSON was identical in both archives, but the PNG files, with embedded workflows, should be different. Anyway, I’ve fixed it.
Regarding “ControlNet,” I set it to “Edit Mode,” uploaded an image to load the image node, selected a preprocessor (I tried depth mask and something else – I don't really remember exactly), entered the prompt, but instead of using something like "change dress color," I used a common prompt in the same way as for txt2img/img2img.
For the inpaint mask, it’s the same as always: right-click on “Load Image” and select “Open in Mask Editor | Mask Canvas.” There's more information within the workflow now, but still don’t expect it to perform as well as SDXL-like models.
Concerning latent editing, it’s similar to working with images – you first need to save your initial .latent file for input. I mean there’s an option “Save Latent with Metadata” and refreshing the workflow (pressing R). I plan to create a more detailed manual with examples and possibly develop my own "Save Latent" node instead of using the core one. For now, just use the standard version of the workflow.