The modeling and quantization method called INT8 ConvRot, which was natively supported successful ComfyUI v0.27.0 released connected July 1, 2026, is simply a basking topic.
It is peculiarly beneficial for the GeForce RTX 20/30 series, but it has besides been reported to supply capacity exceeding the antecedently modular FP8 and FP8 Scaled formats connected the GeForce RTX 40/50 bid arsenic well.
Because of this, it is said that INT8 ConvRot will go the modular for each 8-bit quantized models, and support is really being precocious by Comfy-Org.
Since accusation has not yet dispersed wide and galore group are misunderstanding aliases confusing it, I will shape and explicate it successful this article.
2026/7/7: Added that Forge Neo now supports INT8 ConvRot
2026/7/6: Added accusation regarding Kimama-san's INT8 ConvRot information results
2026/7/5: Added accusation astir GGUF successful "Classification of Modeling Methods and Formats"
2026/7/5: Added accusation astir Triton and PyTorch successful "How to usage INT8 ConvRot"
AI Model Storage Formats
Before explaining INT8 ConvRot, I will explicate the basics of AI exemplary retention formats.
Modeling and Quantization
AI models are modeled successful FP32, a 32-bit FLOAT type, for maximum precision successful a wide range, but since moreover SDXL results successful a ample record size of 12GB, they are converted to trim the spot count to 16-bit, 8-bit, aliases successful immoderate cases 4-bit.
The smaller the spot count, the smaller the record size, making it easier to grip and faster to process. However, because the number of representable values decreases and precision drops, methods and formats to minimize the effect connected precision and results are utilized arsenic overmuch arsenic possible.
Classification of Modeling Methods and Formats
The array beneath astir classifies the modeling methods and formats of AI models.
$$
\begin{array}
{l|l|l}
\bf{階層} & \bf{例} & \bf{役割} \\
\hline
\text{ファイル形式} & \text{safetensors, GGUF,} & \text{保存するコンテナ} \\
& \text{ONNX} & \\
\hline
\text{符号化方式} & \text{FP16, BF16,} & \text{数値をbit列で表現する方法} \\
& \text{INT8, FP8, FP4} & \\
\hline
\text{スケーリング} & \text{なし, Tensor-wise,} & \text{値の範囲を調整する方法} \\
\text{方式} & \text{Row-wise} & \\
\hline
\text{量子化方式} & \text{RTN, AWQ,} & \text{低精度へ変換するアルゴリズム} \\
& \text{GPTQ, ConvRot} & \\
\end{array}
$$
I searched for specified a classification array connected the Web but could not find one, truthful I created it while Q&A-ing pinch ChatGPT for my ain understanding. It mightiness beryllium unsmooth aliases person inappropriate terminology, but since I had it corrected galore times, I judge the wide classification is correct.
2026/7/5 update: GGUF is not conscionable a record format but besides specifies a retention method for quantized information specified arsenic Q4_K_M, truthful it cannot beryllium organized successful the classification array above.
Since this article does not touch connected GGUF, I will instrumentality to the classification successful the array supra truthful arsenic not to complicate the organization.
By the way, successful a wide sense, modeling pinch these methods and formats is sometimes collectively called "quantization," but successful this article, I will telephone it "modeling" and mention to the algorithm for converting to debased precision arsenic a "quantization method."
Note that location are countless sites that explicate individual methods and formats successful detail, truthful I will not explicate them successful this article. If necessary, please look them up elsewhere aliases inquire an AI.
Regarding FLOAT types, Kimama-san's article is detailed.
Representative Model Formats and INT8 ConvRot
Based connected the classification array mentioned above, typical exemplary formats including INT8 ConvRot are organized arsenic follows.
16-bit models
$$
\begin{array}
{l|l|l}
\bf{階層} & \bf{FP16} & \bf{BF16} \\
\hline
\text{符号化方式} & \text{FP16} & \text{BF16} \\
& & \text{(Brain FP16)} \\
\hline
\text{スケーリング} & \text{なし（固定）} & \text{なし（固定）} \\
\text{方式} & & \\
\hline
\text{量子化方式} & \text{RTN} & \text{RTN} \\
\\
\end{array}
$$
8-bit INT type models
$$
\begin{array}
{l|l|l|l}
\bf{Layer} & \bf{INT8} & \bf{INT8} & \bf{INT8} \\
& & \bf{Tensor} \text{-} \bf{wise} & \bf{ConvRot} \\
\hline
\text{Encoding Method} & \text{INT8} & \text{INT8} & \text{INT8} \\
\\
\hline
\text{Scaling} & \text{None (Fixed)} & \text{Tensor-wise} & \text{Row-wise} \\
\text{Method} & & & \\
\hline
\text{Quantization Method} & \text{RTN} & \text{RTN} & \text{ConvRot} \\
\\
\end{array}
$$
8-bit FLOAT model
$$
\begin{array}
{l|l|l|l}
\bf{Layer} & \bf{FP8} & \bf{FP8\ Scaled} & \bf{MXFP8} \\
\hline
\text{Encoding Method} & \text{FP8} & \text{FP8} & \text{FP8} \\
& \text{(E4M3/E5M2)} & \text{(E4M3/E5M2)} & \text{(E4M3/E5M2)} \\
\hline
\text{Scaling} & \text{None (Fixed)} & \text{Tensor-wise, etc.} & \text{Single-layer} \\
\text{Method} & & & \text{Microscaling} \\
\hline
\text{Quantization Method} & \text{RTN} & \text{RTN} & \text{RTN} \\
\\
\end{array}
$$
4-bit model
$$
\begin{array}
{l|l|l}
\bf{Layer} & \bf{NVFP4} & \bf{MXFP4} \\
\hline
\text{Encoding Method} & \text{FP4 (E2M1)} & \text{FP4 (E2M1)} \\
\\
\hline
\text{Scaling} & \text{Layered} & \text{Single-layer} \\
\text{Method} & \text{Microscaling} & \text{Microscaling} & \\
\hline
\text{Quantization Method} & \text{RTN} & \text{RTN} \\
\\
\end{array}
$$
INT8 ConvRot is simply a format modeled utilizing INT8 encoding and a quantization method called ConvRot By comparing it pinch tables of different formats, you should beryllium capable to understand the differences successful methods for each layer.
And conscionable arsenic FP8, FP8 Scaled, and MXFP8 are different, INT8 ConvRot is besides a different format from existing exemplary formats for illustration INT8 aliases INT8 Tensor-Wise. Many group confuse these.
Note that this classification is rough, and location are moreover much elaborate classifications and methods. Multiple formats are besides distributed for INT8 ConvRot. (Described later)
What is INT8 ConvRot?
ConvRot is simply a exemplary format that uses the exertion from the pursuing insubstantial published connected 2025/12/3.
Technically, it tin beryllium mixed pinch FLOAT types arsenic good arsenic INT types, but it is specifically aimed astatine solving the weaknesses of INT types successful AI modeling.
In the first place, INT types (integer types) person a narrower scope of representable values compared to FLOAT types (floating-point types), and location was a problem wherever "outliers" successful AI exemplary processing degraded accuracy. This is why INT8 format models person hardly been utilized until now, and FP8 format has been utilized instead.
In elemental terms, ConvRot is simply a method of Convolution (local artifact processing for illustration convolution) and Rotation (group-based rotation transformation) to distribute outliers crossed aggregate dimensions, converting them into a worth distribution that is not problematic moreover for INT types.
As a result, it has achieved capacity that not only avoids problems but really surpasses FLOAT exemplary formats.
How to usage INT8 ConvRot
INT8 ConvRot tin beryllium utilized successful ComfyUI aliases Forge Neo.
Added connected 2026/7/5: Information is confusing and requirements are unclear, but installing Triton mightiness beryllium basal to get the champion processing speed. It is besides recommended to update Nvidia drivers and PyTorch to the latest versions possible, and it seems amended to usage cu130 alternatively than cu128 for PyTorch.
ComfyUI autochthonal nodes and Comfy-Org charismatic models
ComfyUI has autochthonal support for INT8 ConvRot from v0.27.0 onwards. Additionally, comfy-kitchen 0.2.16 aliases later is required, truthful please update utilizing "pip instal -r requirements.txt" aliases similar. After that, you tin grip INT8 ConvRot format models usually utilizing Load Diffusion Model, Load CLIP, etc.
Regarding models, https://huggingface.co/Comfy-Org/models has added the INT8 ConvRot format to its awesome models complete the past fewer days. As of this writing, I person confirmed that the INT8 ConvRot format has been added to Ideogram-4, Krea-2, Boogu-Image, QIE-2511, SeedVR2, Z-Image-Turbo, Z-Image, Wan_2.2.
Let's usage int8_convrotComfyUI-INT8-Fast and dedicated models
There is besides a measurement to usage a civilization node called ComfyUI-INT8-Fast, which has been released since earlier ComfyUI v0.27.0.
Regarding ComfyUI autochthonal nodes, location were reports that this civilization node was faster owed to issues pinch LoRA successful commits during improvement earlier the v0.27.0 release. However, since this has been addressed successful ComfyUI v0.27.0, we urge utilizing autochthonal nodes unless location is simply a circumstantial request not to.
As a characteristic not recovered successful autochthonal nodes, you tin person models different than the INT8 ConvRot format to the INT8 ConvRot format on-the-fly for use, aliases prevention them straight to a file. For example, if the charismatic root does not administer the INT8 ConvRot format, you tin usage a workflow called int8_save_convrot_model.json to person and create it yourself.
Model format conversion utilizing int8_save_convrot_modelHowever, the INT8 ConvRot format created this measurement is not compatible pinch the autochthonal ComfyUI format, truthful it must beryllium converted utilizing a instrumentality called convert_to_comfy.py.
The aforesaid applies to INT8 ConvRot format models distributed connected HuggingFace for ComfyUI-INT8-Fast.
If you intend to usage them, please publication their respective READMEs carefully.
Forge Neo
Forge Neo did not support INT8 ConvRot astatine the clip this article was written, but it added support for INT8 ConvRot successful perpetrate 6d0bc6e connected 2026/7/7. It will beryllium supported successful the merchandise type 2.27.
No difficult settings aliases operations are required; you tin usage it simply by specifying the ComfyOrg charismatic INT8 ConvRot format exemplary successful the Checkpoint.
Performance of INT8 ConvRot
Hardware support for Nvidia GPUs
The processing capacity for FLOAT and INT types of each spot varies depending connected the GPU model. Basically, newer GPU models person hardware support for much formats and tin process them faster.
The array beneath is simply a database of hardware support for Nvidia GPUs for encoding and scaling methods.
Even if a exemplary format is not supported, it will fundamentally not consequence successful an correction because it is software-converted to a processable format, but it will beryllium slower than erstwhile hardware support is available.
$$
\begin{array}
{l|c|c|c|c}
& \bf{RTX 20} & \bf{RTX 30} & \bf{RTX 40} & \bf{RTX 50} \\
& \bf{Series} & \bf{Series} & \bf{Series} & \bf{Series} \\
\hline
\text{FP16} & \text{✓} & \text{✓} & \text{✓} & \text{✓} \\
\\
\hline
\text{INT8} & \text{✓} & \text{✓} & \text{✓} & \text{✓} \\
\\
\hline
\text{BF16} & & \text{✓} & \text{✓} & \text{✓} \\
\\
\hline
\text{FP8} & & & \text{✓} & \text{✓} \\
\\
\hline
\text{FP8 Scaled} & & & \text{✓} & \text{✓} \\
\\
\hline
\text{MXFP8} & & & & \text{✓} \\
\text{(Microscaling)} \\
\hline
\text{FP4} & & & & \text{✓} \\
\\
\hline
\text{NVFP4} & & & & \text{✓} \\
\text{(Microscaling)} \\
\hline
\text{INT4} & \text{✓} & \text{✓} & \text{✓} \\
\\
\end{array}
$$
FP8, which has been utilized arsenic the modular until now, was only supported by the RTX 40/50 series.
In contrast, INT8 ConvRot is attracting attraction because it tin use from hardware support moreover connected the RTX 20/30 series.
By the way, the FP8 ConvRot format is technically possible, but it is not utilized because it has almost nary advantages successful accuracy aliases capacity compared to INT8 ConvRot.
Regarding 4-bit, the developer of ConvRot seems to person made statements astir INT4 and FP4, but personally, I person ne'er utilized NVFP4, and I would not see utilizing a exemplary pinch specified reduced accuracy for the liking of speed. This is astir apt because I usage an RTX 5090 and seldom consciousness dissatisfied pinch the speed.
Performance reports connected reddit
On reddit, the capacity of INT8 ConvRot is reported arsenic follows.
Processing speed
INT8 ConvRot > NVFP4 > MXFP8 ≒ FP8 Scaled > GGUF Q8 > FP16 ≒ BF16Inference accuracy
FP16 ≒ BF16 > GGUF Q8 > INT8 ConvRot > MXFP8 > FP8 Scaled > FP8 > NVFP4
Originally, 8-bit INT types were accelerated sloppy of accuracy, but INT8 ConvRot moreover outperforms NVFP4 successful velocity connected the RTX 50 series, and its conclusion accuracy approaches that of GGUF Q8.
For the RTX 20/30 series, which lacked hardware support for FP8 and were slow, the benefits are moreover greater than for the RTX 40/50 series.
reddit sources:
https://www.reddit.com/r/StableDiffusion/comments/1tazxqz/
https://www.reddit.com/r/comfyui/comments/1uk6q5m/
I was amazed that INT8 ConvRot is faster than NVFP4, which has a smaller spot count, but aft repeated Q&A pinch ChatGPT, I judge it is owed to the pursuing factors.
ConvRot is simply a processing method optimized for INT types
Performance differences successful hardware and libraries for INT8 and FP4 connected the RTX 50xx
Overhead of NVFP4 Hierarchical Microscaling Processing
Comparison of Krea2 procreation velocity connected existent hardware
I besides measured and compared the image procreation velocity of Krea2 successful my ain situation utilizing the pursuing setup.
GPU: RTX 5090
Driver: Game Ready 610.62 WHQL (2026/6/16)
Python: Python 3.12.12, torch 2.12.1+cu130, triton-windows 3.7.1.post27, sageattention 2.2.0+cu130torch2.9.0andhigher.post4
Soft: ComfyUI v0.27.0 (2026/7/1)
--reserve-lvram 0.9 --preview-method car --use-sage-attentionWF: ComfyUI template Krea-2: Text to Image
prompt_enhance=false
enable_lora?=false
1024 x 1024, 8Steps, cfg=1.0, euler simpleDiT: 5 charismatic Comfy-Org Krea2 Turbo models
TE: qwen3vl_4b_fp8_scaled
VAE: qwen_image_vae
Krea-2: Text to Image$$
\begin{array}
{l|r|r|r}
\bf{DiTモデル} & \bf{DiT, TE, VAE} & \bf{起動直後} & \bf{2回目以降} \\
\bf{形式} & \bf{合計サイズ} & \bf{cold start} & \bf{warm start} \\
\hline
\text{int8\_convrot} & \text{15.24GB} & \text{5.14s} & \text{2.56s} \\
& & \text{(-18\%)} & \text{(-35\%)} \\
\hline
\text{nvfp4} & \text{9.82GB} & \text{5.68s} & \text{3.29s} \\
& & \text{(-9\%)} & \text{(-16\%)} \\
\hline
\text{fp8\_scaled} & \text{14.91GB} & \text{6.25s} & \text{3.91s} \\
\\
\hline
\text{mxfp8} & \text{15.28GB} & \text{6.47s} & \text{4.07s} \\
& & \text{(+4\%)} & \text{(+4\%)} \\
\hline
\text{bf16} & \text{27.15GB} & \text{9.19s} & \text{4.84s} \\
& & \text{(+47\%)} & \text{(+24\%)} \\
\end{array}
$$
"Cold start" is measured aft restarting ComfyUI.
The values successful parentheses are the percentage alteration comparative to fp8_scaled.
My results were accordant pinch the reports connected Reddit. Even connected an RTX 5090, which has hardware support for FP8 Scaled, the procreation clip was
reduced by arsenic overmuch arsenic 35%. For RTX 20/30 bid cards that deficiency hardware support for FP8 Scaled, the simplification should beryllium moreover greater. This is amazing!
I wanted to comparison capacity and generated images erstwhile utilizing LoRA, but I will omit that arsenic this station has go long.
Updated 2026/7/6: The aforementioned article by Kimama-san has been updated pinch information results for INT8 ConvRot procreation clip and accuracy. The accuracy information utilizing generated images is very easy to understand, truthful please mention to it.
Summary
I explained the INT8 ConvRot modeling method and measured and compared the processing speeds. The reports connected Reddit are very affirmative regarding some velocity and accuracy, and while I only measured processing velocity successful my environment, it performed precisely arsenic reported. I americium definite that it will switch FP8 and FP8 Scaled arsenic the modular for 8-bit exemplary formats successful the future.
By the way, it seems immoderate group are converting models for illustration Illustrious aliases Anima—which are already compact and accelerated capable successful FP16 aliases BF16—into INT8 ConvRot, but unless you are highly dissatisfied pinch the speed, I deliberation it is amended to debar doing so.
In this way, I supply easy-to-understand, free explanations of method contented that is not very systematically organized aliases understood. Please support maine pinch a 'Like,' 'Follow,' aliases 'Tip'.
English (US) · 
Indonesian (ID) ·