Workflows, prompts and full test sheets: Skool https://www.skool.com/artfat-ai-6792 · Discord https://discord.gg/WWSe9zSSQ · YouTube https://www.youtube.com/@ArtfatAi
A style LoRA for MOSS Image that takes away the polished stock-photo look and gives you a real phone photo instead. Skin keeps its pores and freckles. Light comes from somewhere you can see. The background stays sharp like on a phone, rooms have their everyday clutter, and people are caught mid-moment instead of posing.
It works on people, couples, groups, food, pets, streets and interiors, in daylight, golden hour and at night, SFW and NSFW.
Trigger: cndamiph at the start of the prompt.
What changes compared to plain MOSS Base (same seed, same prompt)
- stock / editorial look → amateur phone photo
- shallow DSLR blur → deep phone sharpness
- glossy made-up faces → natural skin, freckles, real age
- night scenes keep the lamp or neon light instead of a flat flash
- signs, shirts and notes stay readable
### Recommended setup (used for the gallery)
- MOSS Image Base bf16 + cndamiph 1.0 + an 8-step turbo LoRA by lodestones (Apache-2.0): https://huggingface.co/lodestones/Kromoss
- 8-steps-dmd2-lora — the newest one (out 20 Oct). DMD with an extra GAN stage, gives a sharper, more contrasty picture.
Download: https://huggingface.co/lodestones/Kromoss/resolve/main/8-steps-dmd2-lora.safetensors?download=true
- 8-steps-dmd-lora — the first one, used for this gallery at strength 0.8.
Download: https://huggingface.co/lodestones/Kromoss/resolve/main/8-steps-dmd-lora.safetensors?download=true
- Use one of them, not both. Both work with cndamiph, pick the one you like.
- Difference on the same seed and prompt (both at 0.8): DMD2 is about 14% sharper (cardboard, paper bag and fruit skin show it most). The amateur look stays the same: freckles, side sun, sharp background, no plastic skin. Small background details can come out different. Comparison sheet: 02_grids/14_S1_market_dmd_vs_dmd2.jpg.
- If you skip the SeedVR2 upscale, take DMD2: on close portraits it is about 30% sharper on its own. With SeedVR2 on top the gap shrinks to about 10%, so there both are fine.
- Keep DMD2 at 0.8 too. At 1.0 it changes the face more (eye colour, blush) and adds a little shine to the skin. Close-up sheet: 02_grids/15_portraits_dmd_vs_dmd2_crops.jpg.
- The same seed does not give the same picture with the other turbo LoRA: the composition and the pose shift a little (in NSFW scenes contact and position can change). If you liked a seed on one of them, stay on that one.
- 8 steps · CFG 1 · sampler res_6s · scheduler bong_tangent (RES4LYF)
- optional last pass: downscale ×0.5 → SeedVR2 3B upscale to 2560 px, color correction lab. It adds the fine phone-sharp detail and leaves the picture as it is.
res_6s is a 6-stage sampler, so 8 steps are about 48 model calls. Still far less than a 50-step CFG 3 run, and sharper than plain 8 steps.
### Other modes
| Mode | Steps | CFG | Sampler / scheduler | LoRA strength |
|---|---|---|---|---|
| Fast, plain sampler | 8 + DMD 0.8 | 1 | er_sde / simple | 0.8 |
| Without the speed LoRA | 50 | 3 | er_sde / simple | 0.8 |
| Without the speed LoRA, faster | 25 | 3 | er_sde / simple | 0.8 |
Base model: MOSS Image Base, bf16. Quantized int8 weights noticeably weaken the LoRA, so use bf16 if you can.
### Tips
- Write prompts as plain sentences: who, where, what light, what is lying around.
- Put any text you want in the picture in "double quotes".
- Name the light and its side ("a warm bedside lamp on the right").
- Write "plain unbranded" for clothes and shoes if you do not want logos.
- Without the LoRA loaded, remove the trigger word, or the base model may print it as text.
- On 8 steps, busy group scenes and fine skin come out softer and people can look a few years younger. 50 steps gives the most detail.
- Left / right is not always followed. That comes from the base model.
### Promo videos
The two videos in the gallery were animated with MiniMax H3 straight from this LoRA's own stills.
Music: "Cruising" by Aisake, Dosi and "Bad Habit (feat. Zaug) - Phonk Version" by Zaug, Jéja. Music provided by NoCopyrightSounds (NCS).
### Credits
- 8-step turbo LoRAs 8-steps-dmd-lora, 8-steps-dmd2-lora) by lodestones — https://huggingface.co/lodestones/Kromoss (Apache-2.0). Thanks for making MOSS Image fast.
- Base model: MOSS Image by OpenMOSS — https://huggingface.co/OpenMOSS-Team/MOSS-Image-Base
- Upscale: SeedVR2 · sampler: RES4LYF (res_6s / bong_tangent)
Trained on 80 original images made for this LoRA (60 SFW, 20 NSFW, all adults), captioned in plain prose. MOSS Image Base, rank 32.
All people shown are fictional adults.