Ideogram tip: use Generate Text node to make JSON with Qwen 8B without leaving ComfyUI
https://redd.it/1txmpbi
@rStableDiffusion
https://redd.it/1txmpbi
@rStableDiffusion
Z-Image Turbo vs Flux for AI character consistency — tested both on the same dataset, here's what I found
https://redd.it/1txoucf
@rStableDiffusion
https://redd.it/1txoucf
@rStableDiffusion
Why do half of people hate Ideogram 4.0 and half think it's great?
Each thread about Ideogram 4 seem to have very split comment sections. A lot of people seem to get a frequent censored outputs, find the quality poor, or just find it difficult to use. I've even seen people accuse positive sentiment towards ideogram as astroturfing bots. A lot of other people are praising it for being among the best T2I models currently available for its prompt adherence and image quality.
Using Kijai's prompt builder on the latest stock template worked well for me. Takes a bit of time tweaking the new prompt setup in the builder, but the control it gives makes it worth it for me. I tested a bit of "anatomy" prompting and didn't get any censoring. At 2mp with 3:2 image it took about a minute on a 4090. This model doesn't produce flawless output every time, but it's an improvement to my eyes. The bar for what's "high quality" also seem to go higher and higher, you've probably noticed this if you've been here for a few years.
Where do you fall on Ideogram 4? If you have personally taken the time to test it out a bit, please share. Whether good or bad, I encourage you to share your workflow.
Edit: I really appreciate the discussion in this thread, learning a lot and I can see why both sides have strong feelings for sure
https://redd.it/1txm3tc
@rStableDiffusion
Each thread about Ideogram 4 seem to have very split comment sections. A lot of people seem to get a frequent censored outputs, find the quality poor, or just find it difficult to use. I've even seen people accuse positive sentiment towards ideogram as astroturfing bots. A lot of other people are praising it for being among the best T2I models currently available for its prompt adherence and image quality.
Using Kijai's prompt builder on the latest stock template worked well for me. Takes a bit of time tweaking the new prompt setup in the builder, but the control it gives makes it worth it for me. I tested a bit of "anatomy" prompting and didn't get any censoring. At 2mp with 3:2 image it took about a minute on a 4090. This model doesn't produce flawless output every time, but it's an improvement to my eyes. The bar for what's "high quality" also seem to go higher and higher, you've probably noticed this if you've been here for a few years.
Where do you fall on Ideogram 4? If you have personally taken the time to test it out a bit, please share. Whether good or bad, I encourage you to share your workflow.
Edit: I really appreciate the discussion in this thread, learning a lot and I can see why both sides have strong feelings for sure
https://redd.it/1txm3tc
@rStableDiffusion
Reddit
From the StableDiffusion community on Reddit
Explore this post and more from the StableDiffusion community
Character creation/ design/ manipulation with ZIT and Klein 9B.
https://redd.it/1txs49n
@rStableDiffusion
https://redd.it/1txs49n
@rStableDiffusion
LTX 2.3: You're using it wrong | The Power of Seed Hunting | Workflow in comments
https://www.youtube.com/watch?v=pgV1B3P03D4
https://redd.it/1txtrtr
@rStableDiffusion
https://www.youtube.com/watch?v=pgV1B3P03D4
https://redd.it/1txtrtr
@rStableDiffusion
YouTube
AI Filmmaking Part 1 | Video gen with LTX 2.3: You're doing it wrong.
This is part 1 of a series I am making covering the finer details of AI Filmmaking using 100% local, open source, free models & tools.
Link to workflow (LTX 2.3 Seed Hunting Multiroll):
https://civitai.com/models/2676452
Support me: https://www.patreon…
Link to workflow (LTX 2.3 Seed Hunting Multiroll):
https://civitai.com/models/2676452
Support me: https://www.patreon…
I have trained diffusion and flow matching models from scratch. Same architecture, same dataset, huge difference.
https://redd.it/1ty0vbi
@rStableDiffusion
https://redd.it/1ty0vbi
@rStableDiffusion
Reddit
From the StableDiffusion community on Reddit: I have trained diffusion and flow matching models from scratch. Same architecture…
Explore this post and more from the StableDiffusion community
_____ is the most _____ model you've ever seen!
Why are there hundreds of daily hype posts about >!_____!< model, while everyone here has been sleeping on >!_____!< for days? Have you noticed how you can't say one >!_____!< thing about the quality of >!_____!< in this sub? Meanwhile the bots immediately >!__!<vote every post or comment about >!_____!<!
The people saying that >!_____!< is >!_____!< just don't know how to use >!_____!<. Skill issue! It's not that hard. Step one: use >!_____!< LLM (not >!_____!<!!) to rewrite your prompt. B. use the >!_____!< and >!_____!< custom nodes, and D. stop using >!_____!< and >!_____!< sampler/scheduler!!
### Here's a comparison of both models:
**<cherry picked images|videos>**
How do you idiots not see the difference? There's not a single shred of anything decent or any reason to ever use >!_____!<. It looks exactly like the SD1.5 slop I was making in 2008. Yet it can't even make a young >!_____!< >!_____!<ing upside-down into the mouth of a >!_____!<. What a joke! That company released it because they hate you, they personally forced *me* to install and use it, and the model spits in the face of god!
Meanwhile, the legendary geniuses behind the goated >!_____!< just gave it to us for FREE!! The quality literally obliterates even Seedance. I've yet to see another model that can do a closeup headshot of an attractive young woman. And I guarantee that there will **never** be another model next month that comes even close to being as good. Yet I'm literally the only person who's ever posted about it here.
Oh well, here come the downvotes!
*Sincerely,*
Every top voted post and comment in this sub every day
https://redd.it/1txz2vy
@rStableDiffusion
Why are there hundreds of daily hype posts about >!_____!< model, while everyone here has been sleeping on >!_____!< for days? Have you noticed how you can't say one >!_____!< thing about the quality of >!_____!< in this sub? Meanwhile the bots immediately >!__!<vote every post or comment about >!_____!<!
The people saying that >!_____!< is >!_____!< just don't know how to use >!_____!<. Skill issue! It's not that hard. Step one: use >!_____!< LLM (not >!_____!<!!) to rewrite your prompt. B. use the >!_____!< and >!_____!< custom nodes, and D. stop using >!_____!< and >!_____!< sampler/scheduler!!
### Here's a comparison of both models:
**<cherry picked images|videos>**
How do you idiots not see the difference? There's not a single shred of anything decent or any reason to ever use >!_____!<. It looks exactly like the SD1.5 slop I was making in 2008. Yet it can't even make a young >!_____!< >!_____!<ing upside-down into the mouth of a >!_____!<. What a joke! That company released it because they hate you, they personally forced *me* to install and use it, and the model spits in the face of god!
Meanwhile, the legendary geniuses behind the goated >!_____!< just gave it to us for FREE!! The quality literally obliterates even Seedance. I've yet to see another model that can do a closeup headshot of an attractive young woman. And I guarantee that there will **never** be another model next month that comes even close to being as good. Yet I'm literally the only person who's ever posted about it here.
Oh well, here come the downvotes!
*Sincerely,*
Every top voted post and comment in this sub every day
https://redd.it/1txz2vy
@rStableDiffusion
Reddit
From the StableDiffusion community on Reddit
Explore this post and more from the StableDiffusion community
ComfyUI support or ByteDance Lance-3B (unified image/video generation, editing, and understanding), with dynamic VRAM for low-VRAM GPUs
https://redd.it/1ty4kg1
@rStableDiffusion
https://redd.it/1ty4kg1
@rStableDiffusion
A commercial I recently worked on. AI VFX in the real world!
https://youtu.be/K4ISXPlGx5Y?si=bqh7X1otSWKDghxy
https://redd.it/1ty44d1
@rStableDiffusion
https://youtu.be/K4ISXPlGx5Y?si=bqh7X1otSWKDghxy
https://redd.it/1ty44d1
@rStableDiffusion
YouTube
Make the Shift
If you're only investing in stocks and bonds, it’s time to ask why. Explore alternative strategies with PICTON.
Make the shift: https://www.pictoninvestments.com/en/make-the-shift?utm_source=marketing-social&utm_medium=youtube&utm_content=corporate
Follow…
Make the shift: https://www.pictoninvestments.com/en/make-the-shift?utm_source=marketing-social&utm_medium=youtube&utm_content=corporate
Follow…
Old Man Yells at Node
https://preview.redd.it/efx0wi6mal5h1.png?width=1200&format=png&auto=webp&s=07ef216a5e02f0e2af89a9e1b1e6325a86a7923f
There are a lot of new custom nodes appearing lately. Non-developers, legitimately and rightfully excited about the new superpowers that vibe coding grants them, have begun exploring what they can accomplish. It turns out they can accomplish a lot, because in mid-2026, agentic coding is pretty damn amazing. People who couldn't write a line of code are shipping functional tools.
The thing is, since they're not experienced developers, they aren't thinking about things like maintainability, brittleness, composability, or finding the simplest solution for the task. They just tell Claude to make a thing for them, and Claude does, and it is large and smooth and wonderful, a vibe-coded Jenga tower that sprung fully formed from their mind. And that's fine. The thing works, and the maker is happy and gets some karma and maybe some github stars, and in two weeks nobody ever thinks about the wonderful vibe-coded Jenga tower again.
It is large and smooth and complete. But you're meant to be able to put your hands into a workflow, to stir it up, to affect it. Working the knobs on a sealed box is a legitimate interaction model, but that's what you do with an app. In a workflow, it's kind of a category error.
The vibe-coded Jenga tower is magnificent, but it's also yours, solving your problem your way. Sharing it with me is beside the point because I have the same vibe-coding superpowers as you. I can make my own.
https://redd.it/1ty7cjj
@rStableDiffusion
https://preview.redd.it/efx0wi6mal5h1.png?width=1200&format=png&auto=webp&s=07ef216a5e02f0e2af89a9e1b1e6325a86a7923f
There are a lot of new custom nodes appearing lately. Non-developers, legitimately and rightfully excited about the new superpowers that vibe coding grants them, have begun exploring what they can accomplish. It turns out they can accomplish a lot, because in mid-2026, agentic coding is pretty damn amazing. People who couldn't write a line of code are shipping functional tools.
The thing is, since they're not experienced developers, they aren't thinking about things like maintainability, brittleness, composability, or finding the simplest solution for the task. They just tell Claude to make a thing for them, and Claude does, and it is large and smooth and wonderful, a vibe-coded Jenga tower that sprung fully formed from their mind. And that's fine. The thing works, and the maker is happy and gets some karma and maybe some github stars, and in two weeks nobody ever thinks about the wonderful vibe-coded Jenga tower again.
It is large and smooth and complete. But you're meant to be able to put your hands into a workflow, to stir it up, to affect it. Working the knobs on a sealed box is a legitimate interaction model, but that's what you do with an app. In a workflow, it's kind of a category error.
The vibe-coded Jenga tower is magnificent, but it's also yours, solving your problem your way. Sharing it with me is beside the point because I have the same vibe-coding superpowers as you. I can make my own.
https://redd.it/1ty7cjj
@rStableDiffusion
