This LoRA will turn Ideogram 4 into a few step model. It is capable of generating good quality images in as little as 2 steps with no CFG and no unconditional model. It was created with a continuious turbo training method I am working on and still perfecting. I will release more information on that once it is fully polished. This will likely recieve updated versions as the method improves.
Description
FAQ
Comments (17)
Nice. Was waiting for this.
I managed to bake in trained LoRa to both base models, instead of using two Load Lora nodes which requires heavy background calculations in ComfyUI on the fly. Such pre-merged models reduces time and required resources without any loss. With proper weights, it also eliminates the "safety filter". Baking in Turbo LoRa could be another great addition to both one model and two model modes.
Wow I would love to be able to do that. Do you mind explaining how you were able to do that or where to go for more information?
@faroe_dapol650 I used python, and I don't think I can post longer script codes here. But you can ask one of the larger AI models (Gemini is up to date on ideogram4) to create two files for the purpose of baking a LoRa into the ideogram base models (like merge_ideogram_lora.py and run_bake.bat) for your spesific system and environment.
For weights (in bat file): Lets assume you train a standard character LoRa in Ostris AI-toolkit. In the right lower corner of the session page you'll see a merge button. First, merge the two last saved LoRa checkpoints. This merged character LoRa will require lower weights in ideogram architecture (about 0.4), and prevent the LoRa from overtaking the base model without quality loss. Next, use this weight in the bat file. For the unconditional you might go lower, like 0.3.
If you haven't trained the LoRa yourself, test for the correct weights in a ComfyUI workflow with the standard Lora Load nodes before baking. "Safety filter" outputs means the weight is too low.
This also works with LoKr, but LoRa seems way more stable.
If you run into any troubles, provide the error lines to the AI model. It will guide you and correct coding.
This LoRA doesn't work with other LoRAs though. So does that mean other LoRAs will have to be trained with this LoRA in mind?
Yes I noticed it messes up images and makes to much noise a lot of the time when using with Realism Engine lora as well. Maybe because they are not trained to run at CFG=1? It would be great if someone could fix this, or this lora will have limited uses.
All my images get very dark using other loras. Waiting for an update anxiously.
it was created by ostrich, at least give him credit.
I think you mean Ostris. I checked with him, he said it was cool.
buy GPU or use lora..what do you think who win ? ..
Thanks so much for this Ostris!, Ideogram 4.0 is great, but the 150+ second generation speed was killing me, now I can make a good image in 15 seconds (on my 3090 at 3-4 steps) and maybe seed hunt and upscale to best ones.
Anyone got this working on the gguf version ? i already updated my sd-cli.exe to newest, tried 8 steps also, reults look horrible like 8 steps without lora (no difference).
Maybe only for diffuser model ?
I did some work for that and patched the comfyui gguf loaders.
1. Fetch the patched GGUF loader nodes from here: https://github.com/molbal/ComfyUI-GGUF
2. Get GGUFs from here: https://huggingface.co/molbal/ideogram-4-gguf
3. It works with loading LORAs
What strength should the model be, and do i need to have a lora for both for the uncond and postive ideogram checkpoints?
Good (necessary) initiative, but not ideal results yet. All images get very dark using it when there are other loras along.
This is great für Speed. I really like it and it it perfect for SFW Stuff. On the other hand when it comes to NSFW stuff, it seems to have conflicts and problems showing expplicit details.
how to use this stuff any workflow please?
Is there a workflow for this? I can't see it here or on huggingface.