ArtificialGuyBR

Home / Blog / AI video

Hunyuan Bloom Redmond LoRA: Surreal Floral Video Generation

Generate cinematic surreal floral sculptures with HunyuanVideo using the Bloom Redmond LoRA. Trigger word: Bloom. CC0 license.

2 sources cited AI video

On this page

  1. What This Model Does
  2. Trigger Word & Prompting
  3. How It Was Trained
  4. Setup in ComfyUI
  5. Example Outputs
  6. Recommended Settings
  7. License
  8. Common Errors and Fixes
  9. FAQ
  10. Sources

Hunyuan Bloom Redmond LoRA: Surreal Floral Video Generation

TL;DR

What This Model Does

The Bloom Redmond LoRA transforms HunyuanVideo's output into a distinctive macro-timelapse aesthetic: large-scale, impossible floral sculptures that fuse organic matter, culinary art, entomology, and ceramics. The style draws direct inspiration from Google's Veo/Flow "Bloom" demos — cinematic close-ups where buds unfurl into intricate, layered constructions.

Unlike generic "floral" LoRAs, this one produces structured surrealism: each bloom is a deliberate composition — sushi ingredients, beetle carapaces, porcelain clay — rendered with photorealistic textures and cinematic camera moves (orbit, drift, macro).

Trigger Word & Prompting

The single trigger word is Bloom (case-sensitive in practice). Best results come from prompts that describe a grand, surreal floral sculpture or installation:

Bloom. A macro timelapse opens on a clean, stainless-steel counter. A perfect bud of sushi rice, wrapped in a dark sheet of nori, sits at the center. The bloom begins as the nori unfurls, revealing petals made of paper-thin slices of vibrant pink tuna, orange salmon, and buttery hamachi.

Other verified prompt themes from the model card:

The model responds well to material specificity ("paper-thin slices," "translucent butterfly wings," "throwing rings") and camera direction ("camera drifts inward," "camera orbits," "macro timelapse").

How It Was Trained

The author (ArtificialGuyBR) notes this is a style adaptation of the Bloom aesthetic for HunyuanVideo, not a from-scratch training.

Setup in ComfyUI

  1. Download Hunyuan-Bloom-Redmond.safetensors → ComfyUI/models/loras/
  2. Ensure HunyuanVideo diffusion model (e.g., hunyuan-video-t2v-720p-q4_0.gguf) → ComfyUI/models/diffusion_models/
  3. Place CLIP text encoders → ComfyUI/models/text_encoders/
  4. Place VAE → ComfyUI/models/vae/
  5. Add a Load LoRA node between your base model loader and KSampler
  6. Set strength: 0.8–1.0 (1.0 for full effect)

Works with standard HunyuanVideo ComfyUI workflows. No custom nodes required beyond base HunyuanVideo support.

Sushi bloom example — nori unfurling into tuna, salmon, hamachi petals with avocado, cucumber, tobiko layers and wasabi core

Example Outputs

Prompt Theme | Visual Result

|---|---

Culinary (sushi) | Nori bud opens into layered fish/veg petals; wasabi core in soy sauce pool Entomological | Beetle carapace petals hinge to reveal butterfly wings, cicada layers, glowing firefly center Ceramic | Porcelain clay thrown on wheel → translucent petals with throwing rings → cracking dry texture → cobalt slip core Beetle bloom example — iridescent carapace petals opening to butterfly wings and firefly core

Parameter | Value

|---|---

Base model | tencent/HunyuanVideo (720p) LoRA strength | 0.8–1.0 Resolution | 720p (1280×720 or 720×1280) Sampler | Euler a / DPM++ 2M Karras (standard Hunyuan) Steps | 20–30 CFG | 6–8 Trigger word | Bloom (required)

License

CC0-1.0 (public domain). Free for commercial and non-commercial use, no attribution required. The base HunyuanVideo model has its own license (Apache-2.0 via Tencent).

Common Errors and Fixes

Error | Cause | Fix

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

No style change | LoRA not loaded or strength too low | Verify Load LoRA node connected; set strength ≥ 0.8 "Bloom" ignored | Trigger word misspelled or not at prompt start | Use exactly Bloom. as first token VRAM OOM | 720p HunyuanVideo + LoRA exceeds GPU memory | Use quantized GGUF (q4_0); enable offloading; reduce batch Black/flickering output | VAE or CLIP mismatch | Ensure Hunyuan-matched VAE and CLIP in correct folders

FAQ

Q: Does this work with other video models (WAN, SVD, LTX)? A: No. This LoRA targets HunyuanVideo's architecture specifically.

Q: Can I merge this LoRA into the base model? A: Yes — use ComfyUI's "LoRA Merge" or svd_merge scripts, but keep a backup; merging is destructive.

Q: Is the trigger word mandatory? A: Yes. Without Bloom the style does not activate reliably.

Q: Where can I find more prompts? A: The HF model card includes 4 detailed example prompts; the author's FindGoodPrompt site curates more.

Sources