does it need your Qwen3VL_8b or will any do?
since the model is quite large it would be nice not to have another huge Text Encoder also needed.
Our text encoder is trained, so you need to download it.
since the model is quite large it would be nice not to have another huge Text Encoder also needed.
The qwen3-vl-8b-joyimage_edit_int8_convrot text encoder works also well in ComfyUI with Ideogram4 (but the "Generate Text" is limited to SWF captioning with "Thinking" and "use default template" on). I compared that for Ideogram4 with results from a heretic FP8 text encoder for qwen3-vl-8b and the results are similar but the image generating for Ideogram4 with the int8 text encoder from joyimage_edit is much faster.
I gave up on Joy Image Edit (almost) two weeks ago, but what a difference the Comfy files - and the int8-clip (for my 3050 GPU) make: a great edit model!
thats one of the reasons i asked. Basically anything except as clip, a censored llm model is a waste of time. The babysitting in out of the box models goes well beyond anything that would actually be nsfw or illegal.
also i would argue ideogram doesn't make a great anything, as long as it has these mega cumbersome prompt requirements. with no benefits over a "just do it" editing model. I hope they work on that.
And is fully censored even in itself.
edit: as a sidenote on "NSFW"... i may not use a model "at work". So censoring it to that level of public workplace safety (instead of maybe legal limits) is completely missing the point of a local model. Stop wasting our time with over censoring. .
I agree about censored llm. Problem with the joyimage_edit_int8_convrot text encoder is indeed that it's more censored than Comfy's basic version of the qwen3-vl-8b text encoder, which at least accepts other prompt templates and no thinking for the Text Generate node, the prompt template is crucial (of course that's still no alternative to heretic).
btw i only briefly read up on this, because in comfy templates several in8_convrot turned up as default. apparently int8_convrot is often faster (and better) than fp8 (unlike just int8). That maybe the main reason for your results.
not claiming to understand the technical reasons, it but 5.6 gave me that as general advise.
https://chatgpt.com/share/6a5d1be6-2d1c-83eb-a622-7d0823d7643e
'A useful practical hierarchy reported by the current ComfyUI conversion community is:
Q8 GGUF → INT8 ConvRot → MXFP8 → FP8-scaled → ordinary INT8'
Yes, with my RTX3050 I'm a winner regarding int8_convrot. Past week I was mostly busy searching trough HF's int8 versions for my files. The search on HF isn't exact, neither is Google. For some reason I don't understand I had better luck with duckduckgo. However, out of 10 int8 files two didn't work, despite being Comfy conform, maybe not everything what is uploaded got tested. But the joyimage_edit text encoder is also int8_convrot, as my other one, so that didn't make a difference I think.
'A useful practical hierarchy reported by the current ComfyUI conversion community is:
Q8 GGUF → INT8 ConvRot → MXFP8 → FP8-scaled → ordinary INT8'
This is wrong. GGUF is for Comfy always worse than safetensors files, GGUF , even Q8, doesn't profit from dynamic vram. A Q8 GGUF is never the best option for Comfy. FP32 (32bit!)→BF 16(brainfloating point) →FP16 (16bit, half precision) →INT8+INT8convrot (8bit integer precision) →FP8 (ultra low precision)→MXFP8→ FP8-scaled.
"Ordinary" INT8 isn't worse than any FP8 variant, that's Reddit user's nonsense which Chatbots love to give out as "information". GGUF Q8's quality is similar to FP16 but it takes longer to generate and it makes no more sense to use them with Comfy's dynamic vram optimisation. I use GGUF now only for LM Studio.
if you're purely after quality it may. it may well fit into vram as gguf while f16 does not. And you basically agree that the quality claim is true,. but yeah with a smaller vram (my laptop has 6GB) it makes no sense. and since the whole vram management improving a lot it changed.
edit: i did a little test on the 5090, using the turbo lora, and 6 steps for impatience reasons, the fp8 mixed took (all the same of course) 52 seconds, int8 48s, and Q6 gguf ( i have no q8 anymore) took about 74s. Assuming Q8 is even slower, it is not a great choice. But if quality is the big goal, it is also not completely insane to use. Didn't really run something complicated enough to judge the quality.
And after playing around with boogu, even flux2 dev feels a bit stupid and needs some babysitting, when it comes to getting complicated prompts right. That was always a bit of an issue, may have to do with mistral whatever..
as for fp8 vs int8, it said marginally in quality and may or may not in speed. but it also said with native support it has gotten more compatible. which was my main concern. i assume this smaller, aggressive model is definitely faster. testing later.
"INT8 ConvRot to be somewhere between slightly slower and noticeably faster, depending on the architecture and kernel.
Quality differences to usually be subtle, but ConvRot should have fewer quantization-induced deviations."
And you basically agree that the quality claim is true,.
For Q8 yes. But this isn't a bonus in Comfy, it's now mainly one for LM Studio and other llm apps. And it was good for Comfy users with less VRAM before dynamic VRAM came.
with a smaller vram (my laptop has 6GB) it makes no sense.
I'm on a desktop with 8GB VRAM, but it also matters how much additional system RAM you have (mine has 64 GB RAM). The model card descriptions by the developers only mention a needed VRAM size.
https://blog.comfy.org/p/dynamic-vram-in-comfyui-saving-local
The usage of GGUF files is also dependent on additional RAM: "You don’t need a GPU to run diffusion GGUFs, just a CPU with RAM. VRAM isn’t required but will make inference much faster. For best results, ensure your total usable memory (RAM + VRAM / unified) is slightly larger than the GGUF size; for example, the 4-bit (Q4_K_M) unsloth/Qwen-Image-Edit-2511-GGUF is 13.1 GB, so you should have at least ~13.2 GB of combined memory. " (https://unsloth.ai/docs/blog/comfyui) This is what you have done so far, but I'm sure if you test INT8 ConvRot later, you'll notice a difference.