Batch (folder) image loading

[tttlllrrr]

If one could point "Load Image" at a folder instead of at an image, and cycle through the images as a sequence during a batch output, then you could use frames of an image as controlnet inputs for (batch) img2img restyling, which I think would help with coherence for restyled video frames. That's not possible in Automatic1111.

[AbyszOne]

If one could point "Load Image" at a folder instead of at an image, and cycle through the images as a sequence during a batch output, then you could use frames of an image as controlnet inputs for (batch) img2img restyling, which I think would help with coherence for restyled video frames. That's not possible in Automatic1111.

Yeah. That sucks. An user made a CN batch PR but automatic doesn't allow it. Controlnet extension owner doesn't seems to be interested on add it neither. Ive insisted so many times since day 1 with no success. Im so puzzled why such insanely useful feature is still under the ground.

[joetech]

I think it would be more useful in Comfy, too, since there's you get so much control over the workflow automation. I'd love to see it here.

[ffiras]

I would love to see that too in comfy. batch processing and looping would be so useful

[Nargodian]

It does seem strange that a workflow tool cannot flow work.

[terryxxxxx]

I also really need this feature

[WASasquatch]

There seems to be misconceptions on not only how this node network operates, but how the underlying stable diffusion architecture operates.

You can't give a stable diffusion batch multiple images as inputs. Just one + mask. Then it does X images in a single generation.

The node network is a linear workflow, like most node networks. It's actually special this has the ability to have it auto run the whole network automatically.

What you'd see is a batch load images, and ability to select images to use in node network individually, at once (well sequentially as it goes down the chain).

I mean come on, not even Blender has a load batch images node and loop them and do stuff nodes. Nor substance, etc, etc.

To this effect, I have a Tensor Batch to Image in my WAS Node Suite, where you can select a image to work on, and can use multiple to cover the whole batch. Perhaps I can make a load images node like the one i have now where you can load all images in a directory that is compatible with that node.

As far as comfyui this could be awesome feature to have in the main system (Batches to single image / Load dir as batch of images)

[Nargodian]

I dunno, I just really want to set up a png sequence to automatically loop through them. and be able to reused the workflow from it. Effectively auto queuing all the images in a folder.

[tttlllrrr]

no yeah just like the batch processing in automatic 1111. It's a subtab of img2img, next to inpaint. you just give it an input directory with a lot of images(like video frames) and it applies your prompt to all of them and puts them in an output directory.

[Slarper]

Stop spamming. #23 already has a similar post.

[WASasquatch]

This is already also an established post. You're late, and don't be rude. That post has even derailed about prompts.

[WASasquatch]

To the point on I have made a batch image loaded, it can output either single image by ID relative to count of images, or it can increment the image on each run in ComfyUI. Look out on WAS Node Suite. Will release soon.

Edit: It's updated: https://github.com/WASasquatch/was-node-suite-comfyui

If mode is incremental_image it will increment the images in the path specified, returning a new image each ComfyUI run. If set to single_image it will only return the image relating to the image_id specified.

[M1kep]

Should this be able to be implemented relatively easily now due to #579?

[M1kep]

Basic POC

[M1kep]

Also seems to have some WIP in #707

[WASasquatch]

Yes persistent classes help but is across all sessions (tabs) and not persistent if you needed to continue super large batches like film frames.

[AceGoodwin]

Sorry I'm so late to this party, but I'm still somewhat nubbish to all (well.. most) of this (then again, who isn't eh?).
This would be a Comfy workaround more than a workflow I think.
It seems to me what's lacking mostly is a means of refreshing the graph?
How about instead of trying to force the re-running of the current graph with new images (or prompts, etc), simply push a new graph onto the queue from within a finishing graph.
A Push Node (just a programmatic queue button) that accepts a graph and a trigger image. This trigger would normally be taken from the final output stage of the running graph (when we know we're finished with this graph). Note that the image is only a trigger to a batch handler node to fire off the next round. All the batch handler has to do is replace the appropriate graph settings (including it's own state) and shoot it to the queue when triggered. (Rude n crude I know, but then so was the brute force undo which seems to work just fine).
Of course the danger is inf loops when used stupidly, but sometimes you just can't make omelettes without dynamite.
Whether it's globbing a directory, or piping video frames, as long as the batch handler can pass itself (it's current state specifically) to successive generations correctly, should be safe enough.
Apologies if this idea is incredibly naive.

[osi1880vr]

We have done this in a different tool, im not naming the tool to not get called a spamer making promotion for a different tool, I give the name if you PM me ;)
What we did there was not only allow a loop as such, we created something where you can combine a infinit number of looping nodes into a super loop where each looping node controlls its own loop cycle. This way you can create a loop like run that folder.
or you can do run that folder by looping this and looping that like loop * (loop * loop * loop) * loop
We did this as we had a extra link for the execution of a node, that one we dont have in comfy, but you could just trigger the next iteration by trigger a new entry into the queue, but the magic is here that multiple nodes would be able to controll their looping state and thus their input into the next iteration of the loop.
when thinking batches(loops) think big, dont just think one thing looping, think about a larger loop where multiple loops are interconnected to each other.
the beauty of this is that you can run large experiments while you sleep. I always loved loops since I first implemented one i the other UI I spend time on building it. Once you did the first steps into loops you will also start loving it.
I did stuff like run those prompts with this number of different parameters for the scheduler and for the sampler as well as for what model - lora combination to use. This way you easy create >1000 images with just a few clicks. and so you find the sweet spot of what you try to do while you sleep. To be able to create the same image in all the different combinations , well just use a number of seeds to be used for each outer loop.
And just to take away there scare of complexity in the code, the looping magic is just a very few lines of code inside the looping nodes, and its the same code for each node. in the comfy world all you need to invent on top of the very few lines of loop magic is the trigger next queue entry node at the end of the flow.
to prevent inifinite loops give the trigger next queue entry node a max iterations counter to stop at any given number you like.
Its so easy just do it, I would do it myself, but i can not spend time on this right now, if one like to try get in touch, I show you what I did for the other tool and how easy it can be done.

[AceGoodwin]

Appreciate the response. With your input I think I have a enough clue about roughly how and where it's done. Likely feel good to give my rusty coding skills a workout. Thanks!

[gacela80]

Hi newbie here made this batch face restore workflow. I copy/pasted from different workflows I found.
I think... it works like a charm but maybe can get better. sharing to see if it helps out anybody and please let me know if I am doing this right or if there is any way to improve.
You have to go to extra options and put the number of images in folder to batch count.

[damnmad660]

Thanks for this. as a newbie this really helps me. Loaded it but had an error with the clip text encode prompt node but easily fixed. I'll use this as the basis for a batch upscale workflow.

[JohnBarosanu]

Hey guys, I don't know if you know, but in ComfyUI you have a button named Auto Queue, so when you reach 0 Queue Size, it restarts like a loop. You can access it by pressing Extra Options below the Queue Prompt button. And the Load Image Batch works with no problem. I didn't install something optional for this function. Sorry if I am late with this problem, I just discovered it recently. In the images that I have posted, I showed the settings from my Load Image Batch and the Auto Queue function. Peace!

[subsudo]

That was helpful, thanks!

[menguzat]

any ideas on how to do this with a video? frame by frame i mean.

[alastingepiphany]

The simple and effective way I know of, is to generate a batch of images from an existing video using VLC media player (by enabling the scene video plug-in) it can output frames 1:1 into a new folder and can name the output images by numerical sequence. You can then batch load those images into your comfy UI workflow. once comfy runs the batch, restring the images back into a video.