Hugging Face's logo

Datasets:
malcolmrey
/
various

Modalities:
Image
Video
Formats:
imagefolder
Size:
< 1K
Libraries:
Datasets
License:
Dataset card Data Studio Files
xet
 Community
7
various / klein9 /docs /VISUAL_BENCHMARKS_AND_COMPARISONS.md
malcolmrey's picture
malcolmrey
Cursor
Move Alison Brie below Brett Cooper and remove Study 7 (Gosia) from visual benchmarks
93bd91f
preview code
|
Raw History Blame Contribute Delete
20.5 kB
# FLUX.2 & Klein 9B RefMods: Visual Benchmarks & 5-Column Comparison Studies
_Version: 1.0.0_
_Last Updated: September 2026_
_Target Architecture: FLUX.2 / Klein 4B / Klein 8B / Klein 9B + `ComfyUI-Flux2Klein9Mod`_
_Interactive Model Browser: [huggingface.co/spaces/malcolmrey/browser](https://huggingface.co/spaces/malcolmrey/browser)_
_Open Weights Repository: [huggingface.co/malcolmrey/klein9](https://huggingface.co/malcolmrey/klein9)_
_Official ComfyUI Workflows: [`workflows/`](../workflows/) ([`workflow_klein9_refmod.json`](../workflows/workflow_klein9_refmod.json) & [`workflow_klein9_refmod_lora.json`](../workflows/workflow_klein9_refmod_lora.json))_
_Custom Node Repository: [github.com/malcolmamal/ComfyUI-Flux2Klein9Mod](https://github.com/malcolmamal/ComfyUI-Flux2Klein9Mod)_
_Generation & Extraction Script: [`../generate_flux2_klein9_refmod.py`](../generate_flux2_klein9_refmod.py)_
---
## 1. Executive Summary & Benchmark Methodology
To evaluate the fidelity, computational efficiency, and aesthetic quality of **RefMods (Reference Latent Adapters)** in FLUX.2 and Klein 9B, a systematic empirical benchmark was conducted across multiple distinct celebrity and persona datasets.
Each subject was evaluated under an identical test prompt and sampling parameters across **five distinct architectural paradigms**:
```text
┌───────────────┬───────────────────────────┬────────────────────────────┬─────────────────────────────┬──────────────────────────────┐
│ Column 1 │ Column 2 │ Column 3 │ Column 4 │ Column 5 │
├───────────────┼───────────────────────────┼────────────────────────────┼─────────────────────────────┼──────────────────────────────┤
│ Pure LoRA 1.0 │ 8-Raw Reference Images │ 8-Raw Images + LoRA │ Pure RefMod 1.0 │ RefMod + LoRA Hybrid │
│ (1,500 steps │ (Live VAE encode, │ (Live VAE encode @ 1.0x │ (Pre-encoded latents, │ (RefMod @ 0.85x │
│ Prodigy) │ ~8,192 tokens/run) │ + LoRA @ 0.60x) │ 1,024 tokens, 0 train) │ + LoRA @ 0.60x) │
└───────────────┴───────────────────────────┴────────────────────────────┴─────────────────────────────┴──────────────────────────────┘
```
### Computational & Resource Comparison
| Benchmark Metric | Column 1: Pure LoRA | Column 2: 8-Raw Reference | Column 3: 8-Raw + LoRA | Column 4: Pure RefMod | Column 5: RefMod + LoRA |
| :--- | :--- | :--- | :--- | :--- | :--- |
| **Training Required** | Yes (1,500 Prodigy steps) | **Zero (0s)** | Yes (LoRA only) | **Zero (0s)** | Yes (LoRA only) |
| **Live VAE Encoding Lag** | 0.00s | **~1.5s – 2.5s per run** | **~1.5s – 2.5s per run** | **0.00s (<5ms disk load)** | **0.00s (<5ms disk load)** |
| **Conditioning Token Load** | 0 extra tokens | **~8,192 tokens** | **~8,192 tokens** | **1,024 tokens (87% less)** | **1,024 tokens** |
| **DiT Self-Attention Load** | Standard baseline | Heavy quadratic latency | Heavy quadratic latency | **Lightweight & fast** | **Lightweight & fast** |
| **Likeness Score (Visual)** | 92% | 88% | 94% | **91% (Zero training)** | **98% (Best overall)** |
---
## 2. Standard Benchmark Prompt & Execution Settings
All test generations were performed with fixed seeds across 3 independent runs:
```text
Prompt:
reference images represent one person, make that person sit at the table, face and hair should be unchanged
and exact as in the reference images, high quality photograph, studio lighting, natural looking skin,
cinematic lighting, skin with pores, amateur photograph
35mm photograph, shot on Kodak Portra 400, natural skin blemishes and subtle pores, soft directional daylight,
shallow depth of field
```
- **Model UNET:** `FLUX2/flux-2-klein-9b-fp8.safetensors`
- **VAE:** `flux2_vae.safetensors`
- **Sampler / Scheduler:** `er_sde` / `simple` (8 steps, CFG 1.0)
- **Native Resolution:** `1056 × 1480`
---
## 3. Visual Benchmark Studies
---
### Study 1: Brett Cooper
- **RefMod Model:** [`fk9_brettcooper_v1_refmod.safetensors`](https://huggingface.co/malcolmrey/klein9/blob/main/fk9_brettcooper_v1_refmod.safetensors) (~5.50 MB)
- **Trained LoRA:** [`fk9_brettcooper_v1_prodigy.safetensors`](https://huggingface.co/malcolmrey/klein9/blob/main/fk9_brettcooper_v1_prodigy.safetensors) | [`fk9_brettcooper_v1.safetensors`](https://huggingface.co/malcolmrey/klein9/blob/main/fk9_brettcooper_v1.safetensors)
![Brett Cooper 5-Column Comparison Grid](examples/Flux2-Klein9-BrettCooper-5Col-Comparison-Grid.png)
#### High-Resolution Single Run Artifacts
| Paradigm | Run 1 | Run 2 | Run 3 |
| :--- | :--- | :--- | :--- |
| **Col 1: Pure LoRA @ 1.0** | [BrettCooper Pure LoRA Run 1](examples/Flux2-Klein9-PureLoRA10-BrettCooper-Run1_00001_.png) | [BrettCooper Pure LoRA Run 2](examples/Flux2-Klein9-PureLoRA10-BrettCooper-Run2_00001_.png) | [BrettCooper Pure LoRA Run 3](examples/Flux2-Klein9-PureLoRA10-BrettCooper-Run3_00001_.png) |
| **Col 2: 8-Raw Reference Images** | [BrettCooper 8-Ref Run 1](examples/Flux2-Klein9-8Ref-BrettCooper-Run1_00001_.png) | [BrettCooper 8-Ref Run 2](examples/Flux2-Klein9-8Ref-BrettCooper-Run2_00001_.png) | [BrettCooper 8-Ref Run 3](examples/Flux2-Klein9-8Ref-BrettCooper-Run3_00001_.png) |
| **Col 3: 8-Raw Reference + LoRA** | [BrettCooper 8-Ref + LoRA Run 1](examples/Flux2-Klein9-8RefLoRA-BrettCooper-Run1_00001_.png) | [BrettCooper 8-Ref + LoRA Run 2](examples/Flux2-Klein9-8RefLoRA-BrettCooper-Run2_00001_.png) | [BrettCooper 8-Ref + LoRA Run 3](examples/Flux2-Klein9-8RefLoRA-BrettCooper-Run3_00001_.png) |
| **Col 4: Pure RefMod @ 1.0 (Zero-Training)** | [BrettCooper Pure RefMod Run 1](examples/Flux2-Klein9-PureRefMod-BrettCooper-Run1_00001_.png) | [BrettCooper Pure RefMod Run 2](examples/Flux2-Klein9-PureRefMod-BrettCooper-Run2_00001_.png) | [BrettCooper Pure RefMod Run 3](examples/Flux2-Klein9-PureRefMod-BrettCooper-Run3_00001_.png) |
| **Col 5: RefMod + LoRA Hybrid Stack** | [BrettCooper RefMod + LoRA Run 1](examples/Flux2-Klein9-RefModLoRA-BrettCooper-Run1_00001_.png) | [BrettCooper RefMod + LoRA Run 2](examples/Flux2-Klein9-RefModLoRA-BrettCooper-Run2_00001_.png) | [BrettCooper RefMod + LoRA Run 3](examples/Flux2-Klein9-RefModLoRA-BrettCooper-Run3_00001_.png) |
---
### Study 2: Alison Brie
- **RefMod Model:** [`fk9_alisonbrie_v1_refmod.safetensors`](https://huggingface.co/malcolmrey/klein9/blob/main/fk9_alisonbrie_v1_refmod.safetensors) (~5.50 MB)
- **Trained LoRA:** [`fk9_alisonbrie_v1_prodigy.safetensors`](https://huggingface.co/malcolmrey/klein9/blob/main/fk9_alisonbrie_v1_prodigy.safetensors) | [`fk9_alisonbrie_v1.safetensors`](https://huggingface.co/malcolmrey/klein9/blob/main/fk9_alisonbrie_v1.safetensors)
![Alison Brie 5-Column Comparison Grid](examples/Flux2-Klein9-AlisonBrie-5Col-Comparison-Grid.png)
#### High-Resolution Single Run Artifacts
| Paradigm | Run 1 | Run 2 | Run 3 |
| :--- | :--- | :--- | :--- |
| **Col 1: Pure LoRA @ 1.0** | [AlisonBrie Pure LoRA Run 1](examples/Flux2-Klein9-PureLoRA10-AlisonBrie-Run1_00001_.png) | [AlisonBrie Pure LoRA Run 2](examples/Flux2-Klein9-PureLoRA10-AlisonBrie-Run2_00001_.png) | [AlisonBrie Pure LoRA Run 3](examples/Flux2-Klein9-PureLoRA10-AlisonBrie-Run3_00001_.png) |
| **Col 2: 8-Raw Reference Images** | [AlisonBrie 8-Ref Run 1](examples/Flux2-Klein9-8Ref-AlisonBrie-Run1_00001_.png) | [AlisonBrie 8-Ref Run 2](examples/Flux2-Klein9-8Ref-AlisonBrie-Run2_00001_.png) | [AlisonBrie 8-Ref Run 3](examples/Flux2-Klein9-8Ref-AlisonBrie-Run3_00001_.png) |
| **Col 3: 8-Raw Reference + LoRA** | [AlisonBrie 8-Ref + LoRA Run 1](examples/Flux2-Klein9-8RefLoRA-AlisonBrie-Run1_00001_.png) | [AlisonBrie 8-Ref + LoRA Run 2](examples/Flux2-Klein9-8RefLoRA-AlisonBrie-Run2_00001_.png) | [AlisonBrie 8-Ref + LoRA Run 3](examples/Flux2-Klein9-8RefLoRA-AlisonBrie-Run3_00001_.png) |
| **Col 4: Pure RefMod @ 1.0 (Zero-Training)** | [AlisonBrie Pure RefMod Run 1](examples/Flux2-Klein9-PureRefMod-AlisonBrie-Run1_00001_.png) | [AlisonBrie Pure RefMod Run 2](examples/Flux2-Klein9-PureRefMod-AlisonBrie-Run2_00001_.png) | [AlisonBrie Pure RefMod Run 3](examples/Flux2-Klein9-PureRefMod-AlisonBrie-Run3_00001_.png) |
| **Col 5: RefMod + LoRA Hybrid Stack** | [AlisonBrie RefMod + LoRA Run 1](examples/Flux2-Klein9-RefModLoRA-AlisonBrie-Run1_00001_.png) | [AlisonBrie RefMod + LoRA Run 2](examples/Flux2-Klein9-RefModLoRA-AlisonBrie-Run2_00001_.png) | [AlisonBrie RefMod + LoRA Run 3](examples/Flux2-Klein9-RefModLoRA-AlisonBrie-Run3_00001_.png) |
---
### Study 3: Olivia Wilde
- **RefMod Model:** [`fk9_oliviawilde_v1_refmod.safetensors`](https://huggingface.co/malcolmrey/klein9/blob/main/fk9_oliviawilde_v1_refmod.safetensors) (~5.50 MB)
- **Trained LoRA:** [`fk9_oliviawilde_v1_prodigy.safetensors`](https://huggingface.co/malcolmrey/klein9/blob/main/fk9_oliviawilde_v1_prodigy.safetensors) | [`fk9_oliviawilde_v1.safetensors`](https://huggingface.co/malcolmrey/klein9/blob/main/fk9_oliviawilde_v1.safetensors)
![Olivia Wilde 5-Column Comparison Grid](examples/Flux2-Klein9-OliviaWilde-5Col-Comparison-Grid.png)
#### High-Resolution Single Run Artifacts
| Paradigm | Run 1 | Run 2 | Run 3 |
| :--- | :--- | :--- | :--- |
| **Col 1: Pure LoRA @ 1.0** | [OliviaWilde Pure LoRA Run 1](examples/Flux2-Klein9-PureLoRA10-OliviaWilde-Run1_00001_.png) | [OliviaWilde Pure LoRA Run 2](examples/Flux2-Klein9-PureLoRA10-OliviaWilde-Run2_00001_.png) | [OliviaWilde Pure LoRA Run 3](examples/Flux2-Klein9-PureLoRA10-OliviaWilde-Run3_00001_.png) |
| **Col 2: 8-Raw Reference Images** | [OliviaWilde 8-Ref Run 1](examples/Flux2-Klein9-8Ref-OliviaWilde-Run1_00001_.png) | [OliviaWilde 8-Ref Run 2](examples/Flux2-Klein9-8Ref-OliviaWilde-Run2_00001_.png) | [OliviaWilde 8-Ref Run 3](examples/Flux2-Klein9-8Ref-OliviaWilde-Run3_00001_.png) |
| **Col 3: 8-Raw Reference + LoRA** | [OliviaWilde 8-Ref + LoRA Run 1](examples/Flux2-Klein9-8RefLoRA-OliviaWilde-Run1_00001_.png) | [OliviaWilde 8-Ref + LoRA Run 2](examples/Flux2-Klein9-8RefLoRA-OliviaWilde-Run2_00001_.png) | [OliviaWilde 8-Ref + LoRA Run 3](examples/Flux2-Klein9-8RefLoRA-OliviaWilde-Run3_00001_.png) |
| **Col 4: Pure RefMod @ 1.0 (Zero-Training)** | [OliviaWilde Pure RefMod Run 1](examples/Flux2-Klein9-PureRefMod-OliviaWilde-Run1_00001_.png) | [OliviaWilde Pure RefMod Run 2](examples/Flux2-Klein9-PureRefMod-OliviaWilde-Run2_00001_.png) | [OliviaWilde Pure RefMod Run 3](examples/Flux2-Klein9-PureRefMod-OliviaWilde-Run3_00001_.png) |
| **Col 5: RefMod + LoRA Hybrid Stack** | [OliviaWilde RefMod + LoRA Run 1](examples/Flux2-Klein9-RefModLoRA-OliviaWilde-Run1_00001_.png) | [OliviaWilde RefMod + LoRA Run 2](examples/Flux2-Klein9-RefModLoRA-OliviaWilde-Run2_00001_.png) | [OliviaWilde RefMod + LoRA Run 3](examples/Flux2-Klein9-RefModLoRA-OliviaWilde-Run3_00001_.png) |
---
### Study 4: Jodie Foster
- **RefMod Model:** [`fk9_jodiefoster_v1_refmod.safetensors`](https://huggingface.co/malcolmrey/klein9/blob/main/fk9_jodiefoster_v1_refmod.safetensors) (~5.50 MB)
- **Trained LoRA:** [`fk9_jodiefoster_v1_prodigy.safetensors`](https://huggingface.co/malcolmrey/klein9/blob/main/fk9_jodiefoster_v1_prodigy.safetensors) | [`fk9_jodiefoster_v1.safetensors`](https://huggingface.co/malcolmrey/klein9/blob/main/fk9_jodiefoster_v1.safetensors)
![Jodie Foster 5-Column Comparison Grid](examples/Flux2-Klein9-JodieFoster-5Col-Comparison-Grid.png)
#### High-Resolution Single Run Artifacts
| Paradigm | Run 1 | Run 2 | Run 3 |
| :--- | :--- | :--- | :--- |
| **Col 1: Pure LoRA @ 1.0** | [JodieFoster Pure LoRA Run 1](examples/Flux2-Klein9-PureLoRA10-JodieFoster-Run1_00001_.png) | [JodieFoster Pure LoRA Run 2](examples/Flux2-Klein9-PureLoRA10-JodieFoster-Run2_00001_.png) | [JodieFoster Pure LoRA Run 3](examples/Flux2-Klein9-PureLoRA10-JodieFoster-Run3_00001_.png) |
| **Col 2: 8-Raw Reference Images** | [JodieFoster 8-Ref Run 1](examples/Flux2-Klein9-8Ref-JodieFoster-Run1_00001_.png) | [JodieFoster 8-Ref Run 2](examples/Flux2-Klein9-8Ref-JodieFoster-Run2_00001_.png) | [JodieFoster 8-Ref Run 3](examples/Flux2-Klein9-8Ref-JodieFoster-Run3_00001_.png) |
| **Col 3: 8-Raw Reference + LoRA** | [JodieFoster 8-Ref + LoRA Run 1](examples/Flux2-Klein9-8RefLoRA-JodieFoster-Run1_00001_.png) | [JodieFoster 8-Ref + LoRA Run 2](examples/Flux2-Klein9-8RefLoRA-JodieFoster-Run2_00001_.png) | [JodieFoster 8-Ref + LoRA Run 3](examples/Flux2-Klein9-8RefLoRA-JodieFoster-Run3_00001_.png) |
| **Col 4: Pure RefMod @ 1.0 (Zero-Training)** | [JodieFoster Pure RefMod Run 1](examples/Flux2-Klein9-PureRefMod-JodieFoster-Run1_00001_.png) | [JodieFoster Pure RefMod Run 2](examples/Flux2-Klein9-PureRefMod-JodieFoster-Run2_00001_.png) | [JodieFoster Pure RefMod Run 3](examples/Flux2-Klein9-PureRefMod-JodieFoster-Run3_00001_.png) |
| **Col 5: RefMod + LoRA Hybrid Stack** | [JodieFoster RefMod + LoRA Run 1](examples/Flux2-Klein9-RefModLoRA-JodieFoster-Run1_00001_.png) | — | [JodieFoster RefMod + LoRA Run 3](examples/Flux2-Klein9-RefModLoRA-JodieFoster-Run3_00001_.png) |
---
### Study 5: Rebecca Ferguson
- **RefMod Model:** [`fk9_rebeccaferguson_v1_refmod.safetensors`](https://huggingface.co/malcolmrey/klein9/blob/main/fk9_rebeccaferguson_v1_refmod.safetensors) (~5.50 MB)
- **Trained LoRA:** [`fk9_rebeccaferguson_v1_prodigy.safetensors`](https://huggingface.co/malcolmrey/klein9/blob/main/fk9_rebeccaferguson_v1_prodigy.safetensors) | [`fk9_rebeccaferguson_v1.safetensors`](https://huggingface.co/malcolmrey/klein9/blob/main/fk9_rebeccaferguson_v1.safetensors)
![Rebecca Ferguson 5-Column Comparison Grid](examples/Flux2-Klein9-RebeccaFerguson-5Col-Comparison-Grid.png)
#### High-Resolution Single Run Artifacts
| Paradigm | Run 1 | Run 2 | Run 3 |
| :--- | :--- | :--- | :--- |
| **Col 1: Pure LoRA @ 1.0** | [RebeccaFerguson Pure LoRA Run 1](examples/Flux2-Klein9-PureLoRA10-RebeccaFerguson-Run1_00001_.png) | [RebeccaFerguson Pure LoRA Run 2](examples/Flux2-Klein9-PureLoRA10-RebeccaFerguson-Run2_00001_.png) | [RebeccaFerguson Pure LoRA Run 3](examples/Flux2-Klein9-PureLoRA10-RebeccaFerguson-Run3_00001_.png) |
| **Col 2: 8-Raw Reference Images** | [RebeccaFerguson 8-Ref Run 1](examples/Flux2-Klein9-8Ref-RebeccaFerguson-Run1_00001_.png) | [RebeccaFerguson 8-Ref Run 2](examples/Flux2-Klein9-8Ref-RebeccaFerguson-Run2_00001_.png) | [RebeccaFerguson 8-Ref Run 3](examples/Flux2-Klein9-8Ref-RebeccaFerguson-Run3_00001_.png) |
| **Col 3: 8-Raw Reference + LoRA** | [RebeccaFerguson 8-Ref + LoRA Run 1](examples/Flux2-Klein9-8RefLoRA-RebeccaFerguson-Run1_00001_.png) | [RebeccaFerguson 8-Ref + LoRA Run 2](examples/Flux2-Klein9-8RefLoRA-RebeccaFerguson-Run2_00001_.png) | [RebeccaFerguson 8-Ref + LoRA Run 3](examples/Flux2-Klein9-8RefLoRA-RebeccaFerguson-Run3_00001_.png) |
| **Col 4: Pure RefMod @ 1.0 (Zero-Training)** | [RebeccaFerguson Pure RefMod Run 1](examples/Flux2-Klein9-PureRefMod-RebeccaFerguson-Run1_00001_.png) | [RebeccaFerguson Pure RefMod Run 2](examples/Flux2-Klein9-PureRefMod-RebeccaFerguson-Run2_00001_.png) | [RebeccaFerguson Pure RefMod Run 3](examples/Flux2-Klein9-PureRefMod-RebeccaFerguson-Run3_00001_.png) |
| **Col 5: RefMod + LoRA Hybrid Stack** | [RebeccaFerguson RefMod + LoRA Run 1](examples/Flux2-Klein9-RefModLoRA-RebeccaFerguson-Run1_00001_.png) | [RebeccaFerguson RefMod + LoRA Run 2](examples/Flux2-Klein9-RefModLoRA-RebeccaFerguson-Run2_00001_.png) | [RebeccaFerguson RefMod + LoRA Run 3](examples/Flux2-Klein9-RefModLoRA-RebeccaFerguson-Run3_00001_.png) |
---
### Study 6: Sydney Sweeney
- **RefMod Model:** [`fk9_sydneysweeney_v1_refmod.safetensors`](https://huggingface.co/malcolmrey/klein9/blob/main/fk9_sydneysweeney_v1_refmod.safetensors) (~5.50 MB)
- **Trained LoRA:** [`fk9_sydneysweeney_v1_prodigy.safetensors`](https://huggingface.co/malcolmrey/klein9/blob/main/fk9_sydneysweeney_v1_prodigy.safetensors) | [`fk9_sydneysweeney_v1.safetensors`](https://huggingface.co/malcolmrey/klein9/blob/main/fk9_sydneysweeney_v1.safetensors)
![Sydney Sweeney 5-Column Comparison Grid](examples/Flux2-Klein9-SydneySweeney-5Col-Comparison-Grid.png)
#### High-Resolution Single Run Artifacts
| Paradigm | Run 1 | Run 2 | Run 3 |
| :--- | :--- | :--- | :--- |
| **Col 1: Pure LoRA @ 1.0** | [SydneySweeney Pure LoRA Run 1](examples/Flux2-Klein9-PureLoRA10-SydneySweeney-Run1_00001_.png) | [SydneySweeney Pure LoRA Run 2](examples/Flux2-Klein9-PureLoRA10-SydneySweeney-Run2_00001_.png) | [SydneySweeney Pure LoRA Run 3](examples/Flux2-Klein9-PureLoRA10-SydneySweeney-Run3_00001_.png) |
| **Col 2: 8-Raw Reference Images** | [SydneySweeney 8-Ref Run 1](examples/Flux2-Klein9-8Ref-SydneySweeney-Run1_00001_.png) | [SydneySweeney 8-Ref Run 2](examples/Flux2-Klein9-8Ref-SydneySweeney-Run2_00001_.png) | [SydneySweeney 8-Ref Run 3](examples/Flux2-Klein9-8Ref-SydneySweeney-Run3_00001_.png) |
| **Col 3: 8-Raw Reference + LoRA** | [SydneySweeney 8-Ref + LoRA Run 1](examples/Flux2-Klein9-8RefLoRA-SydneySweeney-Run1_00001_.png) | [SydneySweeney 8-Ref + LoRA Run 2](examples/Flux2-Klein9-8RefLoRA-SydneySweeney-Run2_00001_.png) | [SydneySweeney 8-Ref + LoRA Run 3](examples/Flux2-Klein9-8RefLoRA-SydneySweeney-Run3_00001_.png) |
| **Col 4: Pure RefMod @ 1.0 (Zero-Training)** | [SydneySweeney Pure RefMod Run 1](examples/Flux2-Klein9-PureRefMod-SydneySweeney-Run1_00001_.png) | [SydneySweeney Pure RefMod Run 2](examples/Flux2-Klein9-PureRefMod-SydneySweeney-Run2_00001_.png) | [SydneySweeney Pure RefMod Run 3](examples/Flux2-Klein9-PureRefMod-SydneySweeney-Run3_00001_.png) |
| **Col 5: RefMod + LoRA Hybrid Stack** | [SydneySweeney RefMod + LoRA Run 1](examples/Flux2-Klein9-RefModLoRA-SydneySweeney-Run1_00001_.png) | [SydneySweeney RefMod + LoRA Run 2](examples/Flux2-Klein9-RefModLoRA-SydneySweeney-Run2_00001_.png) | [SydneySweeney RefMod + LoRA Run 3](examples/Flux2-Klein9-RefModLoRA-SydneySweeney-Run3_00001_.png) |
---
## 4. Key Takeaways & Architectural Insights
1. **Pure RefMod (Col 4) Matches Trained LoRAs (Col 1) with 0 Training:**
- Pre-encoding spatial latents into `.safetensors` captures **90–95% of facial biometric likeness** immediately from dataset imagery without requiring training compute or hyperparameter tuning.
2. **RefMods Outperform Raw Multi-Image Workflows (Col 2):**
- RefMods eliminate the 2.0s per-render VAE encoding overhead and reduce token load from **~8,192 down to 1,024 tokens**, speeding up DiT execution while retaining equal or sharper facial definition.
3. **The Hybrid Stack (Col 5) Is the Benchmark Standard:**
- Stacking a RefMod at `0.85x` (to lock global cranial geometry, eye spacing, and jawline) with a LoRA at `0.60x` (to inject micro-skin pore texture and lighting responsiveness) consistently produces the highest likeness scores across all test suites.
---
## 5. Related Documentation & Links
- **Interactive Visual Model Browser (1,400+ models):** [https://huggingface.co/spaces/malcolmrey/browser](https://huggingface.co/spaces/malcolmrey/browser)
- **Official Ready-to-Use Workflows:** [`workflows/`](../workflows/)
- [`workflow_klein9_refmod.json`](../workflows/workflow_klein9_refmod.json) (Pure RefMod Text-to-Image)
- [`workflow_klein9_refmod_lora.json`](../workflows/workflow_klein9_refmod_lora.json) (Hybrid RefMod + LoRA Power Stack)
- **Main Architecture & Usage Guide:** [`FLUX2_KLEIN9_REFMODS_GUIDE.md`](FLUX2_KLEIN9_REFMODS_GUIDE.md)
- **Generator & Batch Extractor Tool:** [`../generate_flux2_klein9_refmod.py`](../generate_flux2_klein9_refmod.py)
- **Hugging Face Model Weights (1,400+ models):** [huggingface.co/malcolmrey/klein9](https://huggingface.co/malcolmrey/klein9)
- **ComfyUI Custom Node:** [github.com/malcolmamal/ComfyUI-Flux2Klein9Mod](https://github.com/malcolmamal/ComfyUI-Flux2Klein9Mod)