Tag: nlnet task

Transparent Generations

Db0Leave a comment

We have another new feature available for people to use from the AI Horde. This is the capacity to use Layer Diffuse to generate images with a transparent background directly (as opposed to stripping the image background with a post-processor).

As someone who’s dabbled into video game development in the past (which was in fact the reason I started the AI Horde) being able to generate sprites, icons and other assets can be quite useful, so once I saw this breakthrough, it immediately became something I wanted to support.

To use this feature, you simply need to flip on the transparent switch if your UI supports it, and the Horde will do the rest. If you’re an integrator, simply send “transparent: true” in your payload.

Take note that the images generated by this feature will not match the image you get with the same seed when transparency is not used! Don’t expect to take an image you like and remove the background this way. For that you need to use the post-processor approach.

Also keep in mind, not every prompt will work well for a transparent image generation. Experiment and find what works for you.

As part of making this update work, me and Tazlin also developed, discovered and fixed a number of other issues and bugs.

What would be most interesting for you is a slight change on how hires-fix works. I discovered that the implementation we were using was using the same amount of steps for the upscaled denoising which was completely unnecessary and wasting compute. So we now use a smart system which dynamically determines how many steps to use for the hires-fix based on the denoising strength you used for hires-fix and the steps for the main generation, and we also exposed a new key on the API where you can directly pass a hires-fix denoising strength.

The second fix is allowing hires-fix on SDXL models, so now you can try to generate larger SDXL images at the optimal resolution.

Finally there were a lot of other minor tweaks and fixes, primarily in the horde-engine. You can read further for more development details on this feature.

This update required a significant amount of work as it required that we onboard a new comfyUI node. Normally this isn’t difficult, but it turns out this node was automatically downloading its own LoRa models on startup, and those were not handled properly for either storage or memory. Due to the efficiency of the AI Horde worker, we do a lot of model preloading along with some fancy footwork in regards to RAM/VRAM usage.

So to make the new nodes work as expected, I had to reach in and modify the methods which were downloading models so that they use our internal mechanisms such as the model manager. Sadly the model manager wasn’t aware of strange models like layer diffuse, so it required me adding a new catch-all class of the model manager for all future utility models like these.

While waiting for Tazlin to be happy with the stability of the code, we discovered another major problem: The face-fixer post-processors we were using until now had started malfunctioning, and generating faces with a weird gray sheen. After some significant troubleshooting and investigation, we discovered that ComfyUI itself on the latest version had switched to a different internal library which didn’t play well with the custom nodes doing the face-fixing.

First I decided to update the code of the face-fixer nodes we were using, which is harder than it sounds, as it also downloads models automatically on startup, which again needs to be handled properly. Updating the custom nodes fixed the codeformer face-fixer, but gfpgan remained broken and the comfyUI devs mentioned that someone would have to fix it. Unfortunately those nodes didn’t seem to be actively maintained anymore so there was little hope to just wait for a quick fix.

Fortunately another custom node developer had run into the same problems, and created a bespoke solution for gfpgan licensed liberally, which I could copy. I love FOSS!

In the meantime, through our usual beta testing process, we discovered that there were still some funkiness in the new hires-fix approach, and Tazlin along with some power users of the community were able to tweak things so that they could work more optimally.

All in all, quite a bit of effort in the past month for this feature, but now we provide something which along with the embedded QR Code generation, I’ve seen very few other GenAI services provide, if at all.

Will you use the new transparent image generation? If so, let us know how! And remember if you have a decent GPU, you can help other generate images by adding your PC onto the horde!

Embedded QR Codes via the AI Horde

Db0

Around the same time last year, the first controlnet for generating QR codes with Stable Diffusion was released I was immediately enamored with the idea and wanted to have it ASAP as an option on the AI Horde. Unfortunately due to a lot of extenuating circumstances [gesticulates wildly] I had neither the time, nor the skills to do it myself, nor the people who could help us onboard it. So this fell on the wayside while way more pressing things were being developed.

Today I’m very excited to announce that I have finally achieved and deployed it to production! QR code generation via the AI Horde is here!

To use is fairly simply, assuming your front-end of choice supports it. You simply provide the text that you want represented as a QR code and the AI Horde will generate a QR code, and then using controlnet, will generate an image where the QR code is embedded into it, as if it’s part of the drawing. You can scan the examples below to see it in action.

You’ll notice that unlike some of the examples you’ll find online elsewhere, the QR code we generate is still fairly noticeable as a QR code, especially when zoomed out, or at a distance. The reason for this is that the more fitting you make to the image, the less likely it is that the QR code is scannable. The implementation I followed to achieve this result is specifically tailored to sacrifice “embedding” for the purpose of scannability.

So when you want to generate QR codes, you need to keep in mind that this is a very finicky workflow. The diffusion process can easily “eat” or modify some components of the QR code so that the final image is not readable anymore. The subject matter and model used matters surprisingly much. Subjects which are somewhat noisy (such as the brain prompt in the featured image above) tend to give enough to the model to work with to reshape that area in a way that creates a QR code. Wheres no matter how hard I tried, I couldn’t get it to generate a QR code with an anime model and an anime woman in the subject.

Along with the basic option to provide the QR Code text, you can also customize some more areas from it. For example you can choose where the QR code will be placed in the image. By default we’ll always display it in the center, but sometimes the composition might be easier if you choose to place it on the side, or to the bottom. You can choose a different prompt for the anchor squares, increase or decrease the border thickness, and more. Your front-end should hopefully be explaining these options to you.

If you want to try and make some yourselves right now, I’ve added the necessary functionality to my Lucid Creations front-end already, so feel free to give it a try right now.

Continue reading further to get some development details.

The road leading to me making this feature available was fairly long. Other than all the other priorities I had for the horde, we also had the misfortune that one of our core contributors on the backend/comfyUI side, went suddenly missing at the end of summer. As I am still more focused the middleware/api and infrastructure (plus so much more, halp!) and Tazlin is focused on efficiency, and code maintenance & quality, we didn’t have the necessary skills to add something as complex as QR code generation.

Once it was clear that our contributor wasn’t coming back and nobody else was stepping up to help, I finally accepted that if I want it done, I have to learn to do that part myself as well. So in the past few months I embarked on a journey to start adding more and more complex comfyUI workflows. First came Stable Cascade which required me to build code which can load 2 different model files at the same time. Then Stable Cascade Remix which required that I wrangle up to 5 source images together.

Note that I’m mostly re-using existing fairly straightforward ComfyUI workflows which do these tasks. I don’t have the bandwidth to learn ComfyUI itself that much. But the work of making said workflows function within the horde-engine with payloads that are send via the AI Horde REST API is quite a complex amount of work on top of those. As I hadn’t built this “translation layer”, I was avoiding that area of the code until now, and this work helped me build up enough knowledge and confidence to be able to pull of translating a much much more complex ComfyUI workflow like the QR codes.

So after many months, I decided it was finally the time to tackle this problem. The first issue is getting an actually good QR Code ComfyUI workflow. Unlike the previous workflows I used, it’s surprisingly difficult to find something that works immediately. Most simple QR Code workflows both required that one generates the QR image externally and generated mostly unscannable images.

I was fortunate enough to run into this excellent devlog by Corey Hanson who not only provided instructions on what works and what doesn’t for QR codes, but even provided a whole repository with prebuilt ComfyUI workflows and a custom node which would also generate a QR code as part of the workflow. Perfect!

Well, almost perfect. Turns out the provided ComfyUI workflows were fairly old, and at the rate GenerativeAI progresses even a couple of months means something can easily be too stale to use. On top of that they were using a lot of extra custom nodes in their examples that didn’t parse, which a ComfyUI newbie like me had to untangle. Finally those workflows were great, especially for local use, but a bit overkill for the horde usage.

So first order of business was to understand, then simplify the workflow to just do the bare needed to get a QR code. Honestly it took me a bit of time to simply get the workflow running in ComfyUI itself and half-way understand what all the nodes were doing. After that I had to translate it to the horde-engine format, which by itself required me to refactor how I parse all comfyUI workflows to make it more maintainable in the future.

Finally QR codes require a lot more potential text inputs, which I didn’t want to start explicitly storing in the DB as new columns as they’re used only for this specific purpose. So I had to come up with a new protocol for sending an open ended amount of extra text values. Fortunately I had already the extra_source_images code deployed so I just copied part of the same logic to speed things up.

And then it was time for unit tests and the public beta and all the potential bugs to fix. Which is when I realized that the results on SD 1.5 models were a bit…sucky, so I went back to ComfyUI itself and actually figured out how to make the workflow work with SDXL as well. The results were way more promising.

Unfortunately while the SDXL QR Codes are way nicer, the requirements to generate them are almost tripled compared to SD 1.5. Not only does one need to run SDXL models, but SDXL controlnets are almost as big as the models themselves. The QR code controlnet is 5G on its own, and all that needs to be loaded in VRAM at the same time as the mode. All this means that even middle-range GPUs struggle to generate SDXL QR codes in a reasonable amount of time. This meant that I also had to adjust the worker to give the option for people serving SDXL models to skip SDXL controlnet, and also properly route this switch via the AI Horde.

Nevertheless, this an areas that makes the AI Horde shine, as those with the necessary power, can support those who need it. Most people will find it really hard or frustrating to generate even a single QR code, never-mind an SDXL one, only to discover that it’s unscannable, but through the horde they can easily generate dozens with very little expertise needed and find the one that works for them.

So It’s been a long journey, but it’s finally here, and the expertise I gained by achieving it also means that I now have enough knowledge to start adding more features via ComfyUI. So stay tuned to see more awesome workflows on the AI Horde!

Image Remix on the AI Horde

Db0

The initial deployment of the Stable Cascade (SC) on the AI Horde supported just text2image workflows, but that was just a subset of what this model can do. We still needed to onboard the rest of its capabilities.

One such capability was the “image variations” option, which allows you to send an image to the model, and get a variation of that image, perhaps with extra stuff added in, using the unClip technology. This required quite a bit of work on hordelib so that it uses a completely different ComfyUI workflow but ultimately this was not so much harder than just adding the img2img capabilities to SC.

The larger difficulty came when I wanted to add the feature to remix multiple images together. The problem being that until now the AI Horde only supported sending a single source image and a single source mask, so a varying amount of images was not possible at all.

So to support this, I needed to touch all areas of the AI Horde. The AI Horde had to accept and upload each of them on my R2 bucket and provide individual download links. The SDK had to know to expect and provide methods to download those images in parallel to avoid delays, to the reGen worker had to be able to receive those images and send them to hordelib which should know how to dynamically adjust a comfyUI pipeline on-the-fly to add as many extra nodes as required.

So after 2 weeks of developing and testing, we finally have this feature available. If your Horde front-end supports the “remix” feature. You can send up to 1-6 images to this workflow along with a prompt, and it will try its best to “squash” them all together into one composition. Note that the more images you send, and the larger the prompt, the harder it will be for the model to “retain” all of them in the composition. But it will try its best.

As an example, here’s how the model remixes my own avatar. You’ll notice that the result can understand the general concepts of the image, but can’t follow it exactly as it’s not doing img2img. The blur is probably caused by the need to upscale my original image, which is something I’d like to fix on the next pass.

Original Avatar
Remix Avatar

Likewise, this is the Haidra logo

Original Logo
Remix Logo

And finally, here’s a remix of both logo and avatar together

Pretty neat, huh?

This ability to send extra source images also lays the groundwork for the Horde to support things like InstantID, which I hope I’ll be able to work on supporting soon enough.

Error Codes and Styling

Db0

Does the above image look scary? If so, you might just just be a software developer!

The above is the result of a long-time coming, but massive pull request to standardize the formatting of the AI Horde code. I’ve been meaning to do this ever since I discovered the black and ruff tools, but I’ve been procrastinating for almost as long. Well, I finally somehow got my ass in motion to do it. Including writing tests, and doing some careful regression testing, It took me like a week in total. And I still didn’t apply all of the ruff checks either.

What this means is that from now on, anyone sending a change, can simply run ruff . --fix && black . and it will automatically format all changes to match our standards. Making the code predictable to read and reducing some bad programming practices and potential tech debt.

Also, as a software dev, finally doing this kind of operation is so satisfying. Not much fun to do, but you’re very happy to have this done. What’s a good analogy for this? a peeling session (post your best analogies in the comments)?

Soon after, I also deployed another change that might be useful for AI Horde integrators out there. I have now added unique error return codes to each error message from the horde. This should make it easier to parse the various errors the horde might spit out with code, instead of having to parse an error message which might potentially change in the future. It also allows you to do things like error code translations (although I think it might be useful to allow people to send translations for the various RCs to the horde as PRs, so that we don’t force every frontend to reinvent them)

I also wrote a README page detailing all the existing RCs.

There’s also been the various bugfixes and improvements on the worker, sdk and hordelib code. Remember to update your reGen worker regularly!

Once again, many thanks to NLNet for providing the funding for such “necessary chore” tasks like these. These kind of things are not a ton of fun to do, as they don’t add any new functionality to the project, but they massively help future development by reducing tech debt.

Webhooks on the AI Horde

Db0

Today I am excited to announce that I have deployed a new feature which allows you to specify a webhook when requesting a generation on the AI Horde. If you do that, once each generation is completed on the AI Horde, it will send POST request to the specified url, with a payload matching the request type.

Apropos, it’s a good time to announce I have started writing some integration information for the AI Horde, which contains information about the available API, and SDKs, and of course, the new webhooks. Feel free to send PRs to improve it!

This new functionality can allow a few more efficient ways of using the AI Horde. For example you could avoid polling the AI Horde every second or so, and rely on webhooks, and only do a manual poll every 30 seconds or so, if the requests have not webhooked over to you yet. The AI Horde will retry a webhook 3 times before giving up, so in case of network issues etc, you can always check the status manually as usual. This approach would reduce the load on the AI horde, while at the same time giving you faster results. It’s what I call a win-win!

Of course, not all clients can support webhooks, so for those who can’t, the existing functionality will continue working as usual.

Ludicrous Speed!

Db0

One very useful feature I’ve been meaning to support for the AI Horde for a while has been request batching. Request batching is the function to generate multiple Stable Diffusion images in parallel, by using internal mechanism to the ML libraries, instead of splitting them into multiple processes. Due to the re-using common parts of the request, it allows the GPU to generate each extra image with just 20% slowdown, instead of 100%, so long as you stay within your GPU’s power.

Soon after we finished adding LCM support, I turned my view to making this a possibility, as between these two features, it could massively increase the overall speed at which the AI Horde completes requests. The only problem is the overall complexity to handling this in the Inference.

Today I’m proud to announce that the AI Horde natively supports smartly batching multiple images in the same request when possible which can result in massive improvements in overall speed! Read on for more details of how we achieved it.

By relying on ComfyUI, the most difficult part was done, and earlier work done by Tazlin and Jug had already prepared the ground to use our hordelib library to handle sending such batched requests to the comfy engine, but I still had a lot of work to do to not only allow the AI Horde accept and queue such loads properly, but also for the worker to be able to understand payloads for multiple images.

Fortunately due to the new setup of the reGen worker, being able to adjust it to accept one job for multiple images and then submit multiple image results at the end was easier than I expected. Of course doing multiple image submissions was the hardest part and I had to basically refactor that whole area of the code.

The AI Horde queuing part was not as code intensive. Making a worker pick up multiple requests when possible was not particularly hard, but not giving the worker more than it can “chew through” was. You see your worker might be able to do 1 image at 2048×2048, and it might be able to do 20 images at 512×512. However give it 20x2048x2048 and it will fall down and die! So this required a bit of fancy footwork. The way I solved this is that the worker declares how many batched images they can do along with its max resolution. The horde then assumes that the worker can safely achieve their max batching at 1/3rd of their max resolution. After this part, as the requested resolution of a job increases, the horde will smartly reduce the amount of batches from a job it will give that worker.

Practically this means that when I declare I can do 20 batches and my max resolution for one image is 2048×2048, then I will pick my full 20 images at 512×512 but will only pick 7 images at my full resolution.

Therefore the AI Horde will continue smartly slicing a request for multiple images into a number of jobs. Only this time instead of each job being 1 image, it can be multiple. Effectively this means that the horde is able to way more efficiently utilize the maximum processing power of each worker and therefore the overall performance improves!

There were a few hiccups along this development as well. For one I realized that the hordelib code did not handle batching for img2img requests at all, so I had to pull up my sleeves and jump into the way hordelib translates requests to comfyUI nodes and figure it all out. It took me a while but now that I understand this better, it will make it easier for me to add even more fancy additions to our comfy workflows!

Another somewhat important problem is that the seed returned by batched requests in comfy is not accurate. The explanation of this is a bit too technical, but at the end of the day, there is extra variable when trying to replicate an image generated via a batch, on top of the generation seed. Currently the horde will return the relevant “batch_id” in the generation metadata, which I hope in the future to use so I can add a way to replicate images from batched requests as well.

For now, if you need to ensure you can always replicate your images via a seed, the best way to do it is to request them using the new disable_batching keyword on your request. Setting this to true will make your request always split to 1 image per job, which is the way the horde used to work until now. However since disable_batching is significantly less optimal than batching, it is only available to trusted users (i.e. those who’ve been running workers for a while) and patreon supporters.

Of course you can continue manually splitting your requests to 1 image per request, but that already has increased kudos costs, and in the future this might get disincentivized further for the health of the AI Horde.

Between batching and LCM proliferation, we’re already starting to see significantly improved generation times on the AI Horde. To the point that with enough priority, you can receive 20x1024x1024 images in less than a minute! A small problem is that currently one of our most popular frontends, Artbot, defaults to manually splitting each request to 1 per image. Nevertheless, Its developer Rockbandit is already hard at work making their requests batching-compatible and once that happens, I expect the overall speed with massively improve!

LCM and multiple versions of LoRas on the AI Horde

Db01 Comment

Finally 2024 is here and this allowed me a bit of free time to work on some of my NLNet tasks. The first thing on my list to tackle was adding LCM support on the AI Horde as it provides massively reduced steps, which for a crowdsourced service like ours, it makes all the difference in how much we can deliver.

For those who don’t know, LCM is a new breakthrough in Stable Diffusion that allows to “finetune” the model in such a way where an image can be generated using 10% of the steps previously required. So an image which would require 30 steps to converge, now needs just 3! That is a massive boost for lower-range GPUs. For high range GPUs, it starts avenues such as video generation as an image can happen at millisecond speeds!

Given the benefits, I wanted to work on this as soon as possible, and given the flexibility of the FOSS GenAI technology enthusiasts, we already had a great way to use LCMs, by using LoRas to turn any SD model into an LCM version.

However there was a snag. You see while the AI Horde already supports all LoRas on CivitAI, we never supported different versions of each, as we never expected anyone would want to use more than the latest. Unfortunately people on CivitAI started using the versioning system as “alternative” versions. And the LCM LoRa was using the same approach, where there was a version for each different sampler.

So the first order of business had to be to allow the AI horde to understand and support all LoRa versions of each LoRa! This took the better part of a full work-week of development and debugging, and then another week of troubleshooting and fixing in beta.

The good news is that this lead to us also identifying and squashing a very frustrating long-running bug where workers would rarely return previous images they’d generated instead of the ones requested. Getting someone else’s image is something we definitely don’t want to ever happen so we’re very happy we figured it out.

With that out of the way, I simply had to update the AI Horde itself to be able to handle the payload for specific LoRa versions, and then add support for the LCM sampler and then some ways to urge users to switch to it.

If you’re an AI Horde integrator, we strongly suggest you change your default settings to utilize LCM LoRas in your generations. You can get them from the same API you receive the model details, under the modelVersions key. To use them, you need to send the exact version ID as a string (found in modelVersions[#]['id]) this won’t accept a version name. You will also need to set is_version: true for the LoRa payload. This will tell the worker to look for a version instead of a LoRa ID.

Sending the LoRa name or ID will continue working as usual, grabbing the latest version (modelVersions[0]) from that list, so you existing implementations should continue working as usual.

Also we recently added AlbedoXL in our model list, to provide a better baseline for SDXL generations than basic SDXL 1.0 which requires a refiner to work. Using Albedo you can get generations that do not require a refiner in your workflow at all and get much less “fuzzy” generations in the process!