CivArchive
    Rainforest Light - Warm Moss
    NSFW
    Preview 143434434
    Preview 143434442
    Preview 143434449
    Preview 143434448
    Preview 143434450
    Preview 143434451
    Preview 143434452
    Preview 143434453
    Preview 143434459
    Preview 143434458
    Preview 143434460
    Preview 143434461
    Preview 143434463
    Preview 143434462
    These are **photographic grade / light LoRAs for Krea 2**, trained on my own Canon raw files from one trip each,
    developed to a specific look before training. They are not subject or character LoRAs and they need **no trigger word**:
    the look is applied to whatever you prompt. Every version was trained on exactly the same photos as its sibling, only the
    grade changed, so the two versions of each model are true opposites you can A/B on one seed.
    
    **Base and settings.** Trained with ai-toolkit on Krea-2-Raw (LoRA rank 32 / alpha 32, lr 1e-4, batch 1, buckets 512/768/1024,
    2,500 steps, flow-matching). Tested on **Krea 2 Turbo FP8** in ComfyUI: 8 steps, CFG 1, `er_sde` / `simple`, 1216×832 and 832×1216,
    text encoder `qwen3vl_4b_fp8_scaled`. Load with a normal LoRA loader on both model and CLIP.
    
    **Strength.** 1.0 is the trained look. 0.5 to 0.7 gives a half-strength grade that still reads as the look. Above 1.2 the
    grade gets heavier and compositions start to drift.
    
    **Prompting.** Plain photographic prompts work best: `a photograph of …` or `a cinematic photograph, …`. No trigger word;
    in fact the token from an earlier build tended to leak into the image as literal text, so these were retrained without one.
    
    **img2img.** The Moab and Kauai versions work as colour grades on your own photos at denoise 0.45–0.55 (structure kept,
    grade applied). The Olympic Peninsula versions are place-and-light LoRAs rather than grades and barely change a sunny
    source below denoise 0.65; see their notes.
    
    **What they do to off-topic prompts.** All six leave the composition of unrelated subjects alone (portraits, food, city
    streets, wildlife) and only shift tone, with one exception noted on the Olympic Peninsula page: those two override the
    prompt's lighting toward overcast daylight.
    
    All showcase images are one seed rendered twice, LoRA on (top) and LoRA off (bottom), same prompt, so what you see
    is exactly what the LoRA adds.

    Description

    Brightest of my six and the lowest contrast, with olive-green foliage and lifted blacks. It relights scenes to soft
    overcast daylight and pulls the palette toward warm moss and grey sky.
    
    - **Best for:** elven canopy cities, overgrown wrecks, rangers and wolves in old-growth forest, cabins and canoes in the
      rain, anything Pacific Northwest. Also good for turning any forest prompt from "concept-art dark" into "documentary".
    - **Overrides light:** yes. A twilight alpine village and a dusk lighthouse both became overcast afternoon on the same
      seed. Do not use it on prompts that need sun or night.
    - **Quirks:** adds content of its own: junipers into deserts, a red boat behind a fisherman, a dirt path through a
      lavender field. Skin goes slightly olive. As an img2img grade on sunny photos it is weak below denoise 0.65.
    - **Strength:** 0.6–1.2 all clean; 1.2 pushes foliage toward pale yellow-olive.

    FAQ

    Comments (2)

    hushichoSep 21, 2026
    CivitAI

    Wonderful!

    fantaseedSep 21, 2026
    CivitAI

    You are writing all your explanations inside code blocks</>, but browser translation tools cannot translate the content within code blocks into one's native language; therefore, if possible, please avoid using code blocks for standard explanatory text.

    LORA
    Krea 2

    Details

    Downloads
    122
    Platform
    CivitAI
    Platform Status
    Available
    Created
    9/21/2026
    Updated
    9/25/2026
    Deleted
    -

    Files

    olympic-peninsula-warm-moss-2026_2500.safetensors