Choosing Your HiDream AI Model: Versions, Formats, Hardware Needs & Download Links
HiDream now provided in several AI model versions and formats (like .safetensors and .gguf), each suited for different needs regarding speed, quality, VRAM usage, and software compatibility. Understanding these helps you select the best option for your hardware and find the right download links.
This guide covers:
The core models (Full, Dev, Fast) and where to find them.
The required Text Encoder and VAE models.
The quantized .gguf model versions (Full, Dev, Fast) and their sources.
General GPU VRAM requirements for these models.
1. Core HiDream Image Models
HiDream Full
Focus: Highest Quality & Detail
Settings: Steps: 50, Sampler: uni_pc, Scheduler: simple, CFG: 5.0 (Needs Negative Prompt)
HiDream Dev
Focus: Speed & Experimentation
Settings: Steps: 28, Sampler: lcm, Scheduler: normal, CFG: 1.0 (No Negative Prompt)
HiDream Fast
Focus: Maximum Speed
Settings: Steps: 16, Sampler: lcm, Scheduler: normal, CFG: 1.0 (No Negative Prompt)
Download Location (Split FP16/FP8 Diffusion Models):
https://huggingface.co/Comfy-Org/HiDream-I1_ComfyUI/tree/main/split_files/diffusion_models
Note: Download all parts for the desired model (e.g., hidream_i1_full_fp16.safetensors) and place them in the ComfyUI/models/diffusion_models/ folder.
2. Text Encoder & VAE Models
These components are essential for interpreting your prompts and finalizing the image. They need to be loaded alongside the main image models, consuming additional VRAM.
Required HiDream Text Encoders:
clip_g_hidream.safetensors (1.39 GB)
clip_l_hidream.safetensors (248 MB)
llama_3.1_8b_instruct_fp8_scaled.safetensors (9.08 GB)
t5xxl_fp8_e4m3fn_scaled.safetensors (5.16 GB)
Download Location (Split Text Encoders):
https://huggingface.co/Comfy-Org/HiDream-I1_ComfyUI/tree/main/split_files/text_encoders
Note: Download all parts for the required encoders (clip_g, clip_l, t5xxl ,llama3.1) and place them in the ComfyUI/models/text_encoders/ folder.
VAE (Variational Autoencoder):
HiDream uses the same VAE as the FLUX models
3. GGUF Model Versions (Quantized HiDream)
Available HiDream-I1 GGUF Quantizations & Download Links:
HiDream Full GGUF: https://huggingface.co/city96/HiDream-I1-Full-gguf
Includes various quantizations (FP16, Q8_0, Q6_K, Q5_K_M, Q5_K_S, Q4_K_M, Q4_K_S, Q4_0, Q3_K_M, Q3_K_S, Q2_K) ranging from 34.2 GB (FP16) down to 6.56 GB (Q2_K).
HiDream Fast GGUF: https://huggingface.co/city96/HiDream-I1-Fast-gguf
Provides quantized versions specifically for the faster "Fast" model architecture. Check the repository for available quantization levels and sizes.
HiDream Dev GGUF: https://huggingface.co/city96/HiDream-I1-Dev-gguf
Provides quantized versions specifically for the "Dev" model architecture. Check the repository for available quantization levels and sizes.
Note: K-quants (_K_M, KS) often offer better quality than older standard quants (_0, _1) at similar bit levels.
4. GPU VRAM Requirements & Model Suitability
VRAM is crucial. You need enough space for the model, necessary auxiliaries (encoders, VAE), software overhead, and the generation process.
General GGUF Suitability by GPU VRAM (Focusing on HiDream Full GGUF sizes):
8GB VRAM: Best suited for Q2_K (6.56 GB) or potentially Q3_K_S (8.21 GB) / Q3_K_M (8.77 GB) with some offloading.
12GB VRAM (e.g., RTX 3060 12GB): Can comfortably run Q3 levels fully in VRAM. Can run Q4 levels (up to Q4_K_M @ 11.5 GB), but the larger Q4 models will likely require some layers offloaded to CPU/RAM for smooth operation. Q5 levels (up to 13.5 GB) are possible but require more significant offloading. Recommendation: Start with Q4_K_M or Q4_K_S; try Q3_K_M if speed is an issue due to offloading.
16GB VRAM: Can likely run Q5 levels (up to 13.5 GB) fully in VRAM. Can run Q6_K (14.7 GB), possibly needing minimal offload.
24GB VRAM (or more): Can comfortably run Q6_K (14.7 GB) and likely higher quantizations like Q8_0. The full FP16 (34.2 GB) requires substantial offloading even on 24GB, better suited for 48GB+ cards or multi-GPU setups if aiming for full VRAM utilization.
Conclusion
HiDream offers flexibility:
Use the .safetensors / split files from the Comfy-Org repo for standard GPU workflows (like ComfyUI), ensuring you also download the corresponding text encoders and the FLUX VAE. Best for users comfortable managing these components.
Use the .gguf files from the city96 repos for broader compatibility (LM Studio, etc.) and better handling of VRAM limitations via quantization and CPU offloading. Simpler setup, often just needing one file.
Match the model format and quantization level (for GGUF) to your GPU's VRAM and intended software for the best experience.
I'm using this model (Q4_K_M-dev)(12G-16G) u can drop any image there for workflow
https://civitai.com/models/1479706
You can find NF4 & gguf big collections @RalFinger profile
https://civitai.com/models/1457126/hidream-i1-full-dev-fast-nf4