📸 Krea 2 MistyCharm (FP8 Optimized)
A state-of-the-art DiT checkpoint fine-tuned to deliver raw photographic authenticity, organic micro-textures, complex light transport, and authentic social media aesthetics without the artificial "AI look".
Mathematically balanced on CPU with intact FP8 weights for ultra-fast sampling convergence and production-grade stability.
⚡ Core Strengths & Capabilities
Organic Skin Textures (No Plastic Finish): Completely removes the waxy, over-smoothed appearance common in diffusion models. Retains visible pores, fine blemishes, natural peach fuzz, nail ridges, and expression micro-wrinkles.
Anatomical Integrity & Hand Grips: Resolves complex finger joints, tendon tension, and physical object interactions naturally (smartphones, chisel tools, glassware, coffee cups).
Social Media & Influencer Aesthetics: Excels at creating ultra-convincing digital influencers in casual smartphone photos, realistic mirror selfies featuring flash glares and dust smudges, and spontaneous outdoor street snaps (candid OOTD).
Dynamic Lighting, Volumetrics & Refraction: Handles dramatic chiaroscuro, sharp rim lights, curling volumetric smoke, and authentic wet-surface reflections without edge artifacts or blown-out highlights.
Fast Sampling Convergence: Reaches full catalog-grade detail in just 10 to 14 steps.
Zero Negative Prompt Dependency: Its advanced language understanding parses plain natural language directly without requiring long walls of negative tags.
⚙️ Recommended Generation Settings
ParameterOptimal ValueSupported RangeSampler / KSampler
Euler
Euler / Euler AScheduler
Beta (fastest) or Simple
Beta / Simple
Sampling Steps
12 – 14 steps
10 to 20 steps (10 with Beta)
CFG Scale
1.0 – 1.8
1.0 to 2.0 (avoid CFG > 2.5)
Guidance
3.5
3.0 to 4.0Negative Prompt
(Leave Empty)
Optional (minimal tags only)
Base Resolutions
768x1152 (portrait) or 1152x768 (landscape)
Standard native aspect ratios
🧪 Validated Stress Tests & Showcase
Macro Handcraft & Tool Physics: Detailed wood carving, visible sawdust, realistic muscle strain, and nail edge textures.
Cinematic Night Scapes: 65mm anamorphic style wet alleys, authentic neon typography, and puddle reflections under translucent umbrellas.
Stage & High-Contrast Light: Live stage musicians with golden rim lights, realistic sweat sheen, and curling smoke paths.
Fine Art & Boudoir Studies: Anatomically accurate torso forms, natural body creases, and delicate sheer lace textile detail.
Smartphone Mirror Selfies: Hand-held phone camera snaps with authentic glass reflections, phone flash bursts, and casual room backdrops.
💡 Prompting Advice
Use descriptive natural language: Mention camera equipment (e.g., shot on smartphone, 35mm lens), specific lighting setups (soft morning window light, dappled sunlight), and casual human expressions.
Anchor the environment: Populate the setting with everyday objects (unmade bed, wooden table, paper coffee cup, vintage street backdrop) to ground the generation in reality.
Keep CFG modest: Keep CFG between 1.0 and 1.5 to maintain maximum dynamic range in the shadow tones and avoid over-contrasted edges.
## ✦ 𝐖𝐡𝐲 𝐈𝐍𝐓𝟖 𝐒𝐜𝐚𝐥𝐞𝐝 (𝐂𝐨𝐦𝐟𝐲𝐔𝐈 𝐍𝐚𝐭𝐢𝐯𝐞)?
Unlike standard raw truncation, this checkpoint uses **dynamic per-tensor scaling weight_scale + comfy_quant)**. By mapping the weight matrices across signed 8-bit integers with high-precision float32 scale factors, it preserves dynamic range and color depth remarkably well:
⟢ *Sharp Pore & Sensor Detail:** Avoids the "waxy/smooth" degradation often seen in basic low-bit conversions.
⟢ *Low VRAM Footprint (~11–13 GB):** Fits comfortably on RTX 3060, 3070, 3080, and 4060/4070 cards without heavy system RAM offloading.
⟢ *Full Turbo Speed:** Retains the exact 8–10 step responsiveness of the Kroma Turbo architecture.
## 📦 𝐀𝐯𝐚𝐢𝐥𝐚𝐛𝐥𝐞 𝐕𝐞𝐫𝐬𝐢𝐨𝐧𝐬 & 𝐏𝐫𝐞𝐦𝐢𝐮𝐦 𝐀𝐜𝐜𝐞𝐬𝐬
⚡ *INT8 Scaled (Free Download Here):** Best balance of low VRAM usage and dynamic range for ComfyUI users.
🚀 *FP8 Edition (Free Download Here):** Fast floating-point alternative for modern Ada/Hopper architectures.
💎 *BF16 Master 24GB (Premium / Exclusive):** Full uncompressed mathematical precision. Available exclusively via our Cloud Studio & Gumroad!
## ⚙️ 𝐑𝐞𝐜𝐨𝐦𝐦𝐞𝐧𝐝𝐞𝐝 𝐒𝐞𝐭𝐭𝐢𝐧𝐠𝐬 (𝐂𝐨𝐦𝐟𝐲𝐔𝐈)
| Parameter | Recommended Value | Notes |
| :--- | :--- | :--- |
| **Loader Node** | Load Diffusion Model | Native ComfyUI support comfy_quant / int8) |
| **Sampler** | Euler | Crisp, stable renders with zero artifacts |
| **Scheduler** | Simple or Normal | Best gradient transitions on Krea 2 / DiT |
| **Steps** | 8 - 12 | 10 steps is the sweet spot for raw skin texture |
| **Guidance / CFG** | 2.0 - 3.5 | 2.5 keeps lighting natural without contrast burn |
| **Resolution** | 832 x 1216 / 1024 x 1024 | Native aspect ratios recommended |
## 💡 𝐄𝐱𝐚𝐦𝐩𝐥𝐞 𝐏𝐫𝐨𝐦𝐩𝐭𝐬
📸 *Raw Sensor Portrait:**
> raw photo, a stunning 24yo latina woman, warm bronze undertones, natural skin texture with visible pores, subtle freckles, soft wavy dark hair, wearing a fitted beige linen top, golden hour sunlight filtering through balcony, 85mm lens, f/1.8, shallow depth of field, photorealistic, unretouched
🌙 *Candid Night / Street Bokeh:**
> candid street photography, young woman smiling softly at the camera, wearing an off-shoulder black sweater, warm neon city lights blurred in the background, 35mm film grain, lifelike skin tone, sharp eye focus, natural evening atmosphere
## 🛡️ 𝐎𝐩𝐭𝐢𝐨𝐧𝐚𝐥 𝐍𝐞𝐠𝐚𝐭𝐢𝐯𝐞 𝐏𝐫𝐨𝐦𝐩𝐭
```text
plastic skin, airbrushed, 3d render, cgi, doll, smooth filter, overexposed, bad anatomy, deformed hands, blurry, watermark
### ⚠️ FAQ / Troubleshooting: "Wrong Checkpoint Loader" Error on INT8 Version
If you get a "Wrong checkpoint loader" error when loading the quantized INT8 version, it means ComfyUI's standard Load Checkpoint node doesn't recognize 8-bit quantized weights natively.
#### 🔧 How to Fix It (Local ComfyUI):
1. Use the correct Loader Node:
* Do NOT use the default Load Checkpoint or Pixaroma nodes.
Use Load Quantized Checkpoint** or *Unet Loader** (requires ComfyUI-GGUF or ComfyUI-BitsAndBytes nodes depending on your setup).
2. Setup your CLIP & VAE:
* Load your text encoders (CLIP) and VAE separately if using a split workflow.
---
### 💡 Tired of INT8 Loader Bugs or Quality Loss?
Get the Uncompressed 24GB BF16 Master Edition!
✅ *100% Plug-and-Play:** Loads natively with standard ComfyUI Load Diffusion Model / Load Checkpoint nodes.
✅ *Zero Quantization Artifacts:** Maximum skin detail, micro-textures, and full dynamic range.
⚡ *High-Speed Cloud Generation / Direct Download:** Available via our Gumroad VIP Vault.
👉 [Get the Uncompressed 24GB BF16 Master Checkpoint Here](https://charmingcontours.gumroad.com/l/gbobrd)
Description
Many of you requested an ultra-efficient integer-quantized version that runs smoothly on consumer GPUs (8GB–12GB VRAM) without sacrificing the signature skin micro-textures, natural lighting, and anatomical accuracy of the model.
FAQ
Comments (17)
BF16 please
The 24GB BF16 Master Checkpoint (uncompressed raw weights for maximum skin texture fidelity and 100% native ComfyUI compatibility) is available on our Gumroad VIP Vault:
Get the 24GB BF16 Master Edition Here
Once unlocked, you get instant high-speed direct download access via our Hugging Face VIP Space! Thanks for the support!
@Charmingcontours No thanks. NEVER
Error.
Tell me what error you get
When start run show 'error'. Wrong checkpoint loader. I try different loaders - Pixaroma, regular... same.
@Charmingcontours im also getting an error
[ERROR] !!! Exception during processing !!! Expecting value: line 1 column 1 (char 0)
[ERROR] Traceback (most recent call last):
File "D:\ComfyUI-Easy-Install\ComfyUI-Easy-Install\ComfyUI\execution.py", line 548, in execute
output_data, output_ui, has_subgraph, has_pending_tasks = await get_output_data(prompt_id, unique_id, obj, input_data_all, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, v3_data=v3_data)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "D:\ComfyUI-Easy-Install\ComfyUI-Easy-Install\ComfyUI\execution.py", line 353, in get_output_data
return_values = await _async_map_node_over_list(prompt_id, unique_id, obj, input_data_all, obj.FUNCTION, allow_interrupt=True, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, v3_data=v3_data)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "D:\ComfyUI-Easy-Install\ComfyUI-Easy-Install\ComfyUI\custom_nodes\comfyui-lora-manager\py\metadata_collector\metadata_hook.py", line 177, in async_map_node_over_list_with_metadata
results = await original_map_node_over_list(
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "D:\ComfyUI-Easy-Install\ComfyUI-Easy-Install\ComfyUI\execution.py", line 327, in _async_map_node_over_list
await process_inputs(input_dict, i)
File "D:\ComfyUI-Easy-Install\ComfyUI-Easy-Install\ComfyUI\execution.py", line 315, in process_inputs
result = f(**inputs)
^^^^^^^^^^^
File "D:\ComfyUI-Easy-Install\ComfyUI-Easy-Install\ComfyUI\nodes.py", line 1014, in load_unet
model = comfy.sd.load_diffusion_model(unet_path, model_options=model_options)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "D:\ComfyUI-Easy-Install\ComfyUI-Easy-Install\ComfyUI\comfy\sd.py", line 2445, in load_diffusion_model
model = load_diffusion_model_state_dict(sd, model_options=model_options, metadata=metadata, disable_dynamic=disable_dynamic)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "D:\ComfyUI-Easy-Install\ComfyUI-Easy-Install\ComfyUI\comfy\sd.py", line 2437, in load_diffusion_model_state_dict
model.load_model_weights(new_sd, "", assign=model_patcher.is_dynamic())
File "D:\ComfyUI-Easy-Install\ComfyUI-Easy-Install\ComfyUI\comfy\model_base.py", line 373, in load_model_weights
m, u = self.diffusion_model.load_state_dict(to_load, strict=False, assign=assign)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "D:\ComfyUI-Easy-Install\ComfyUI-Easy-Install\python_embeded\Lib\site-packages\torch\nn\modules\module.py", line 2609, in load_state_dict
load(self, state_dict)
File "D:\ComfyUI-Easy-Install\ComfyUI-Easy-Install\python_embeded\Lib\site-packages\torch\nn\modules\module.py", line 2597, in load
load(child, child_state_dict, child_prefix) # noqa: F821
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "D:\ComfyUI-Easy-Install\ComfyUI-Easy-Install\python_embeded\Lib\site-packages\torch\nn\modules\module.py", line 2580, in load
module._load_from_state_dict(
File "D:\ComfyUI-Easy-Install\ComfyUI-Easy-Install\ComfyUI\comfy\ops.py", line 1442, in _load_from_state_dict
_load_quantized_module(self, super()._load_from_state_dict, *args, load_extra_params=True)
File "D:\ComfyUI-Easy-Install\ComfyUI-Easy-Install\ComfyUI\comfy\ops.py", line 1244, in _load_quantized_module
layer_conf = json.loads(layer_conf.numpy().tobytes())
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "json\__init__.py", line 346, in loads
File "json\decoder.py", line 338, in decode
File "json\decoder.py", line 356, in raw_decode
json.decoder.JSONDecodeError: Expecting value: line 1 column 1 (char 0)@lowkeylayers same in forge neo
Yes, the INT8 version does not seem to be correct.
If you upload the BF16, people can convert it to the version they want.
If you don't want to upload the BF16, I would recommend silveroxides convert_to_quant project on GitHub 👍
@herkus_baronas631 Of course! That Wrong checkpoint loader error when moving from BF16 to INT8 in ComfyUI happens because the models quantized to INT8 (such as files that lower compression or are processed with 8-bit schemas) no longer fit in the standard load node (Load Checkpoint / Load Diffusion Model).
When compressing to INT8, the structure of the layers changes and the classic loader gets confused or rejects the file looking for the original architecture (such as Krea 2 or FLUX in BF16/FP16).
To fix that error and make it load perfectly, do the following:
Do not use the normal Load Checkpoint or Pixaroma's. Depending on the custom nodes you have installed, you must use:
Load Quantized Checkpoint (if you use quantization nodes such as ComfyUI-GGUF or 8-bit quantization suites).
UNETLoader or the specific ComfyUI-BitsAndBytes / GPTQ if the model is saved in compressed 8-bit format.
@PsychOut The INT8 version works perfectly on ComfyUI using quantized load nodes such as Load Quantized Checkpoint (requires the extension ComfyUI-GGUF or BitsAndBytes). You have the detailed steps at the end of the description.
If you prefer not to install additional nodes and use the ComfyUI native charger without weight compression, the 24GB BF16 Master version is available in our Gumroad.
@lowkeylayers ### 🔧 Fix for json.decoder.JSONDecodeError: Expecting value on INT8
If you get this JSON decoding error when executing the workflow, it means ComfyUI's native UNET loader load_unet) is trying to automatically read non-existent quantization JSON metadata from the checkpoint.
To solve this:
1. Do NOT use the standard *Load Diffusion Model** or *UNET Loader** node with the INT8 file.
2. Use *Load Quantized Checkpoint** (from ComfyUI-GGUF or ComfyUI-BitsAndBytes nodes) which handles non-standard 8-bit tensors without looking for missing JSON metadata.
3. Make sure your .safetensors file downloaded completely (verify file size) and update your ComfyUI to the latest version.
(Note: If you want to bypass quantization errors and custom node requirements entirely, the *Uncompressed 24GB BF16 Master Edition** loads seamlessly with standard nodes).*
@Charmingcontours everyone elses model works fine its on you
@Charmingcontours Sorry, I don't use Comfy, so I can't use the custom loaders necessary.
In the future, maybe silveroxides convert_to_quant project on GitHub will help you create INT8 versions that doesn't need these custom loaders, so that more people can try your model without getting errors, in both Comfy and in Forge.
But thank you for responding! 👍
new model int8 is not even loading I dont even get anything?
Can confirm Int8 in Comfy tosses an error.
INT8 cant be downloaded?