The acceleration with sage+torchcompile on Z-Image is really good.
https://redd.it/1pjswpl
@rStableDiffusion
https://redd.it/1pjswpl
@rStableDiffusion
Reddit
From the StableDiffusion community on Reddit: The acceleration with sage+torchcompile on Z-Image is really good.
Explore this post and more from the StableDiffusion community
ComfyUI-LoaderUtils Load Model When It Need
Hello, I am **xiaozhijason** aka **lrzjason**. I created a helper nodes which could load any models in any place of your workflow.
# 🔥 The Problem Nobody Talks About
~~ComfyUI’s native loader has a dirty secret:~~ **~~it loads EVERY model into VRAM at once~~** ~~– even models unused in your current workflow. This wastes precious memory and causes crashes for anyone with <12GB VRAM. No amount of workflow optimization helps if your GPU chokes before execution even starts.~~
**Edit: Model loads into RAM rather VRAM and dynamic load it when used. So, it doesn't load all models into VRAM at once which is incorrect in the statement.**
# ✨ Enter ComfyUI-LoaderUtils: Load Models Only When Needed
I created a set of **drop-in replacement loader nodes** that give you **precise control over VRAM usage**. How? By adding a magical optional `any` parameter to every loader – letting you **sequence model loading** based on your workflow’s actual needs
https://preview.redd.it/46s01mhwck6g1.png?width=2141&format=png&auto=webp&s=ccfac9d74d1a970cb97e821db6ae6a8b3c211261
**Key innovation:**
✅ **Strategic Loading Order** – Trigger heavy models (UNET/Diffusion model) *after* text encoding
✅ **Zero Workflow Changes** – Works with existing setups (just swap standard loaders for `_Any` versions and connect the loader before it need)
✅ **All Loaders Covered:** Checkpoints, LoRAs, ControlNets, VAEs, CLIP, GLIGEN – \[full list below\]
# 💡 Real Workflow Example (Before vs After)
**Before (Native ComfyUI):**
`[Checkpoint] + [VAE] + [ControlNet]` → **LOAD ALL AT ONCE** → 💥 *VRAM OOM CRASH*
**After (LoaderUtils):**
1. Run text prompts & conditioning
2. *Then* load UNET via `UNETLoader_Any`
3. *Finally* load VAE via `VAELoader_Any` after sampling → **Stable execution on 8GB GPUs** ✅
# 🧩 Available Loader Nodes (All _Any Suffix)
|Standard Loader|Smart Replacement|
|:-|:-|
|||
|`CheckpointLoader`|→ `CheckpointLoader_Any`|
|`VAELoader`|→ `VAELoader_Any`|
|`LoraLoader`|→ `LoraLoader_Any`|
|`ControlNetLoader`|→ `ControlNetLoader_Any`|
|`CLIPLoader`|→ `CLIPLoader_Any`|
|*(+7 more including Diffusers, unCLIP, GLIGEN, etc.)*||
**No trade-offs:** All original parameters preserved – just add connections to the `any` input to control loading sequence!
https://redd.it/1pjvn14
@rStableDiffusion
Hello, I am **xiaozhijason** aka **lrzjason**. I created a helper nodes which could load any models in any place of your workflow.
# 🔥 The Problem Nobody Talks About
~~ComfyUI’s native loader has a dirty secret:~~ **~~it loads EVERY model into VRAM at once~~** ~~– even models unused in your current workflow. This wastes precious memory and causes crashes for anyone with <12GB VRAM. No amount of workflow optimization helps if your GPU chokes before execution even starts.~~
**Edit: Model loads into RAM rather VRAM and dynamic load it when used. So, it doesn't load all models into VRAM at once which is incorrect in the statement.**
# ✨ Enter ComfyUI-LoaderUtils: Load Models Only When Needed
I created a set of **drop-in replacement loader nodes** that give you **precise control over VRAM usage**. How? By adding a magical optional `any` parameter to every loader – letting you **sequence model loading** based on your workflow’s actual needs
https://preview.redd.it/46s01mhwck6g1.png?width=2141&format=png&auto=webp&s=ccfac9d74d1a970cb97e821db6ae6a8b3c211261
**Key innovation:**
✅ **Strategic Loading Order** – Trigger heavy models (UNET/Diffusion model) *after* text encoding
✅ **Zero Workflow Changes** – Works with existing setups (just swap standard loaders for `_Any` versions and connect the loader before it need)
✅ **All Loaders Covered:** Checkpoints, LoRAs, ControlNets, VAEs, CLIP, GLIGEN – \[full list below\]
# 💡 Real Workflow Example (Before vs After)
**Before (Native ComfyUI):**
`[Checkpoint] + [VAE] + [ControlNet]` → **LOAD ALL AT ONCE** → 💥 *VRAM OOM CRASH*
**After (LoaderUtils):**
1. Run text prompts & conditioning
2. *Then* load UNET via `UNETLoader_Any`
3. *Finally* load VAE via `VAELoader_Any` after sampling → **Stable execution on 8GB GPUs** ✅
# 🧩 Available Loader Nodes (All _Any Suffix)
|Standard Loader|Smart Replacement|
|:-|:-|
|||
|`CheckpointLoader`|→ `CheckpointLoader_Any`|
|`VAELoader`|→ `VAELoader_Any`|
|`LoraLoader`|→ `LoraLoader_Any`|
|`ControlNetLoader`|→ `ControlNetLoader_Any`|
|`CLIPLoader`|→ `CLIPLoader_Any`|
|*(+7 more including Diffusers, unCLIP, GLIGEN, etc.)*||
**No trade-offs:** All original parameters preserved – just add connections to the `any` input to control loading sequence!
https://redd.it/1pjvn14
@rStableDiffusion
"Outrage" Short AI Animation (Wan22 I2V ComfyUI)
https://youtu.be/-HeVTeniWv8
https://redd.it/1pjzs92
@rStableDiffusion
https://youtu.be/-HeVTeniWv8
https://redd.it/1pjzs92
@rStableDiffusion
YouTube
177 | "Outrage" | Short AI Animation (Wan22 I2V ComfyUI) [4K]
"Outrage" Short AI Animation
Input Images - Flux1
Video - Wan 2.2 14b I2V + VACE Clip joiner + Wan 2.2 creative upscale, via ComfyUI
100% AI Generated with local open source models
____________________________________________
Let me know your feedback…
Input Images - Flux1
Video - Wan 2.2 14b I2V + VACE Clip joiner + Wan 2.2 creative upscale, via ComfyUI
100% AI Generated with local open source models
____________________________________________
Let me know your feedback…
Z-Image first generation time
Hi, I'm using ComfyUI/Z-image with a 3060 (12GB VRAM) and 16 GB RAM. Anytime I change my prompt, the first generation takes between 250-350 seconds, but subsequent generations for the same prompt are must faster, around 25-60 seconds.
Is there a way to reduce the generation of the first picture to be equally short? Since others haven't posted this, is it something with my machine? (Not enough RAM, etc?)
https://redd.it/1pk13tx
@rStableDiffusion
Hi, I'm using ComfyUI/Z-image with a 3060 (12GB VRAM) and 16 GB RAM. Anytime I change my prompt, the first generation takes between 250-350 seconds, but subsequent generations for the same prompt are must faster, around 25-60 seconds.
Is there a way to reduce the generation of the first picture to be equally short? Since others haven't posted this, is it something with my machine? (Not enough RAM, etc?)
https://redd.it/1pk13tx
@rStableDiffusion
Reddit
From the StableDiffusion community on Reddit
Explore this post and more from the StableDiffusion community
Old footage upscale/restoration, how to? Seedvr2 doesn't work for old footage
https://redd.it/1pk4m9m
@rStableDiffusion
https://redd.it/1pk4m9m
@rStableDiffusion