Freedom Redmond: High-Resolution SD 2.1 Checkpoint

TL;DR

• Generalist SD 2.1 768 checkpoint optimized for resolutions above 1024px
• Handles realistic, artistic, vehicle, and portrait generation with consistent quality
• Recommended: DPM++ 2M SDE Karras, 30 steps, CFG 6, no negative prompt needed in many cases

What This Model Does

Freedom.Redmond is a fine-tuned checkpoint based on Stable Diffusion 2.1 768, trained with GPU support from Redmond.AI. Unlike the base SD 2.1 which natively outputs 768×768, this model is specifically tuned to generate coherent images at 1024px and higher without the common artifacts (repeating patterns, structural collapse) that plague naive upscaling.

The model is a generalist — it produces:

• Photorealistic scenes and portraits
• Artistic illustrations
• Vehicles (cars prominently featured in samples)
• People and characters

It was designed to serve as a foundation for future specialized LoRAs, meaning the base checkpoint captures broad capability rather than a narrow style.

Training & Technical Details

Property | Value

|---|---

Base model | Stable Diffusion 2.1 768 Training focus | High-resolution generation (>1024px) Format | fp16 SafeTensor File size | 2.43 GB License | BigScience OpenRAIL-M Commercial use | Allowed with attribution (@artificialguybr) Verified on Civitai | June 10, 2023 The Civitai page tags include: base model, anime, woman, female, photorealistic, illustration — confirming the versatile output range.

Recommended Generation Settings

From the author's testing:

• Sampler: DPM++ SDE / 2M or 2M SDE Karras
• Steps: 30
• CFG Scale: 6
• Resolution: Above 768px (1024px+ recommended)
• Negative prompt: Optional — some prompts work better without negatives

These settings come directly from the model card. The DPM++ 2M SDE Karras combination balances speed and quality at high resolution.

Example Outputs

The sample images confirm the model's range: one shows a photorealistic portrait with natural skin rendering and lighting, while another demonstrates clean artistic illustration with strong composition. Both hold up at high resolution without the tiling artifacts common in base SD 2.1 at >768px.

How to Use

Automatic1111 / WebUI: Place FreedomFinal.safetensors in models/Stable-diffusion/, select it in the checkpoint dropdown, apply recommended settings.

ComfyUI: Load via CheckpointLoaderSimple node. The fp16 SafeTensor loads natively.

Diffusers (Python):

from diffusers import StableDiffusionPipeline
import torch

pipe = StableDiffusionPipeline.from_single_file(
    "FreedomFinal.safetensors",
    torch_dtype=torch.float16
).to("cuda")

image = pipe(
    "your prompt here",
    num_inference_steps=30,
    guidance_scale=6,
    height=1024,
    width=1024
).images[0]

Common Errors and Fixes

Error | Cause | Fix | Source

|---|---|---|---

Repeating patterns at 1024px+ | Using base SD 2.1 settings | Use this fine-tune; keep resolution ≥1024px | Model card OOM on 2.43 GB model | VRAM exhaustion | Enable xformers, use fp16, reduce batch size | General SD practice Washed-out colors | CFG too low | Raise CFG to 6–7 | Author recommendation

FAQ

Q: Does this require a specific VAE? A: No specific VAE mentioned; the bundled fp16 SafeTensor includes the VAE. The base SD 2.1 VAE works.

Q: Can I use negative prompts? A: Yes, but the author notes many prompts work better without negatives. Test both.

Q: Is this an anime model? A: It's a generalist. Civitai tags include "anime" but samples show photorealistic and illustration styles — not exclusively anime.

Q: What license applies? A: BigScience OpenRAIL-M. Commercial use permitted with attribution to @artificialguybr.

Sources

• HuggingFace model card: https://huggingface.co/artificialguybr/freedom
• Civitai model page: https://civitai.com/models/87288

More models from ArtificialGuyBR: | SD 2.1 checkpoints: | High-res generation guide: