Floating Heads HiDream

The Floating Heads HiDream LoRA is LyCORIS-based and trained on stylized, human-focused 3D bust renders. I had an idea to train on this trending prompt I spotted on the Sora explore page. The intent is to isolate the head and neck with precise framing, natural accessories, detailed facial structures, and soft studio lighting.

Results are 1760x2264 when using the workflow embedded in the first image of the gallery. The workflow is prioritizing visual richness, consistency, and quality over mass output.

That said outputs are generally very clean, sharp and detailed with consistent character placement, and predictable lighting behavior. This is best used for expressive character design, editorial assets, or any project that benefits from high quality facial renders. Perfect for img2vid, LivePortrait or lip syncing.

Workflow Notes

The first image in the gallery includes an embedded multi-pass workflow that uses multiple schedulers and samplers in sequence to maximize facial structure, accessory clarity, and texture fidelity. Every image in the gallery was generated using this process. While the LoRA wasn’t explicitly trained around this workflow, I developed both the model and the multi-pass approach in parallel, so I haven’t tested it extensively in a single-pass setup. The CFG in the final pass is set to 2, this gives crisper details and more defined qualities like wrinkles and pores, if your outputs look overly sharp set CFG to 1.

The process is not fast — expect 300 seconds of diffusion for all 3 passes on an RTX 4090 (sometimes the second pass is enough detail). I'm still exploring methods of cutting inference time down, you're more than welcome to adjust whatever settings to achieve your desired results. Please share your settings in the comments for others to try if you figure something out.

I don't need you to tell me this is slow, expect it to be slow (300 seconds for all 3 passes).

Trigger Words:

h3adfl0at, 3D floating head

Recommended Strength: 0.5–0.6

Recommended Shift: 5.0–6.0

Version Notes

v1: Training focused on isolated, neck-up renders across varied ages, facial structures, and ethnicities. Good subject diversity (age, ethnicity, and gender range) with consistent style.

v2 (in progress): I plan on incorporating results from v1 into v2 to foster more consistency.

Training Specs

• Trained for 3,000 steps, 2 repeats at 2e-4 using SimpleTuner (took around 3 hours)

• Dataset of 71 generated synthetic images at 1024x1024

• Training and inference completed on RTX 4090 24GB

• Captioning via Joy Caption Batch 128 tokens

I trained this LoRA with HiDream Full using SimpleTuner and ran inference in ComfyUI using the HiDream Dev model.

If you appreciate the quality or want to support future LoRAs like this, you can contribute here:
🔗 https://ko-fi.com/renderartist
🔗 renderartist.com