Update to Ideogram4 JSON prompt tool
https://preview.redd.it/dgj9neords5h1.png?width=1364&format=png&auto=webp&s=c9f252cfea67f6b5d37218621ef75306d604b641
A couple days ago I made a small JSON prompt builder for IdeogramV4 which some people found helpful, so posting a small update here.
Main changes are PNG metadata import, photo/art\_style mode switching, and larger image size ranges.
# PNG import
You can drag an image generated in Comfy into the tool's window, and if it contains a JSON prompt in the correct format, it will load the bounding boxes and JSON parameters used to create the image. Imported images do not need to be made with this tool, any image generated with ComfyUI with a valid JSON prompt should work.
This should help with quicker iteration
# Photo/art_style mode switching
The [official Ideogram4 prompting guide](https://github.com/ideogram-oss/ideogram4/blob/main/docs/prompting.md) lists two modes for JSON prompt structure, one for photographic and one for non-photographic images. Both are now supported in the tool.
(I HIGHLY recommend reading the prompting guide, there are lots of useful tips to get the most out of this model)
# Image size range
Ideogram 4 performs surprisingly well with large images, so the image size slider ranges have been increased up to 4096.
The github is [https://github.com/d-daley/ideogram4-editor](https://github.com/d-daley/ideogram4-editor) if you want to clone the tool locally, or you can access it in the browser here: [https://d-daley.github.io/ideogram4-editor/](https://d-daley.github.io/ideogram4-editor/)
https://redd.it/1tz1uaq
@rStableDiffusion
https://preview.redd.it/dgj9neords5h1.png?width=1364&format=png&auto=webp&s=c9f252cfea67f6b5d37218621ef75306d604b641
A couple days ago I made a small JSON prompt builder for IdeogramV4 which some people found helpful, so posting a small update here.
Main changes are PNG metadata import, photo/art\_style mode switching, and larger image size ranges.
# PNG import
You can drag an image generated in Comfy into the tool's window, and if it contains a JSON prompt in the correct format, it will load the bounding boxes and JSON parameters used to create the image. Imported images do not need to be made with this tool, any image generated with ComfyUI with a valid JSON prompt should work.
This should help with quicker iteration
# Photo/art_style mode switching
The [official Ideogram4 prompting guide](https://github.com/ideogram-oss/ideogram4/blob/main/docs/prompting.md) lists two modes for JSON prompt structure, one for photographic and one for non-photographic images. Both are now supported in the tool.
(I HIGHLY recommend reading the prompting guide, there are lots of useful tips to get the most out of this model)
# Image size range
Ideogram 4 performs surprisingly well with large images, so the image size slider ranges have been increased up to 4096.
The github is [https://github.com/d-daley/ideogram4-editor](https://github.com/d-daley/ideogram4-editor) if you want to clone the tool locally, or you can access it in the browser here: [https://d-daley.github.io/ideogram4-editor/](https://d-daley.github.io/ideogram4-editor/)
https://redd.it/1tz1uaq
@rStableDiffusion
Ideogram 4: a solution for removing the annoying censorship has been found.
Actually, there are two working methods. In both cases, the block message almost never appears.
Method 1
Shift only the first sigma step by +0.005.
In some cases, +0.01 may be needed.
This slightly moves the starting point of the trajectory away from what the model expects. All other sigma steps should remain exactly the same as in the default settings.
https://preview.redd.it/sn5kmzicys5h1.png?width=1668&format=png&auto=webp&s=c8f014e508eb8e6cda7024aa107be163f8b7a8f1
Method 2 — preferred
Increase the initial noise by x2.
In some cases, x3 may be needed.
The effect is similar: it pushes the model away from its usual starting trajectory.
https://preview.redd.it/43vp15rfys5h1.png?width=2171&format=png&auto=webp&s=46a885178417fce64273de7eceec2ce91425fdab
Important
Both methods work properly only with an LCM sampler.
Based on the behavior, LCM seems able to correct the trajectory after the initial deviation, so the image still converges properly.
Other samplers will break the image.
Required nodes:
Noise Math from the More Math addon
SamplerLCMCustom from Extra Samplers
Custom Sigmas from KJ Nodes
Sigmas2 Mult from RES4LYF
These hacks remove the gray censorship square, but they do not fix all of the model’s behavior issues.
Ideogram 4 does not seem to be well-trained for very short prompts. Even if one-word prompts no longer trigger censorship, the model may still fail to follow them accurately.
Prompts made of several sentences already work much better, even without JSON formatting.
For this part of the workflow, I use two samplers sequentially. The first sampler runs the initial high-sigma stage with LCM and the modified noise/sigma setup. Then its output is passed into the second sampler, which continues the remaining steps with the normal sigma range.
https://preview.redd.it/3m7dy3yuzs5h1.png?width=2048&format=png&auto=webp&s=b2e41ce79d7a9d12c373495bf77bbdaad2d0c8f0
The split point in SplitSigmas can be adjusted depending on the situation. In my tests, values between 1 and 3 usually work best.
This hack works — at least for me 🙂
But I still think we should keep looking for other ways to get rid of this censorship.
Happy generating, everyone.
https://redd.it/1tz4fnf
@rStableDiffusion
Actually, there are two working methods. In both cases, the block message almost never appears.
Method 1
Shift only the first sigma step by +0.005.
In some cases, +0.01 may be needed.
This slightly moves the starting point of the trajectory away from what the model expects. All other sigma steps should remain exactly the same as in the default settings.
https://preview.redd.it/sn5kmzicys5h1.png?width=1668&format=png&auto=webp&s=c8f014e508eb8e6cda7024aa107be163f8b7a8f1
Method 2 — preferred
Increase the initial noise by x2.
In some cases, x3 may be needed.
The effect is similar: it pushes the model away from its usual starting trajectory.
https://preview.redd.it/43vp15rfys5h1.png?width=2171&format=png&auto=webp&s=46a885178417fce64273de7eceec2ce91425fdab
Important
Both methods work properly only with an LCM sampler.
Based on the behavior, LCM seems able to correct the trajectory after the initial deviation, so the image still converges properly.
Other samplers will break the image.
Required nodes:
Noise Math from the More Math addon
SamplerLCMCustom from Extra Samplers
Custom Sigmas from KJ Nodes
Sigmas2 Mult from RES4LYF
These hacks remove the gray censorship square, but they do not fix all of the model’s behavior issues.
Ideogram 4 does not seem to be well-trained for very short prompts. Even if one-word prompts no longer trigger censorship, the model may still fail to follow them accurately.
Prompts made of several sentences already work much better, even without JSON formatting.
For this part of the workflow, I use two samplers sequentially. The first sampler runs the initial high-sigma stage with LCM and the modified noise/sigma setup. Then its output is passed into the second sampler, which continues the remaining steps with the normal sigma range.
https://preview.redd.it/3m7dy3yuzs5h1.png?width=2048&format=png&auto=webp&s=b2e41ce79d7a9d12c373495bf77bbdaad2d0c8f0
The split point in SplitSigmas can be adjusted depending on the situation. In my tests, values between 1 and 3 usually work best.
This hack works — at least for me 🙂
But I still think we should keep looking for other ways to get rid of this censorship.
Happy generating, everyone.
https://redd.it/1tz4fnf
@rStableDiffusion
Uncensoring Ideogram 4: Whats the Consensus, Whats the best Practice(s)? Discuss Here :)
People have posted some amazing Ideogram 4 gens. The models been out for a while. What a lot of people want to know however is best practices to follow with this model for the least censorship.
Please upvote to get more interactions
https://redd.it/1tz2qkj
@rStableDiffusion
People have posted some amazing Ideogram 4 gens. The models been out for a while. What a lot of people want to know however is best practices to follow with this model for the least censorship.
Please upvote to get more interactions
https://redd.it/1tz2qkj
@rStableDiffusion
Reddit
From the StableDiffusion community on Reddit
Explore this post and more from the StableDiffusion community