rockyPHM — Krea 2 LoRA

A character LoRA for Rocky from Andy Weir's Project Hail Mary.

The character belongs to Andy Weir. The visual design in this LoRA is my own interpretation — it is not an official or canonical depiction.

Status: work in progress. It works well enough to be useful, but it has real limitations (listed below) and I'm still expanding the dataset. Feedback and failure cases are very welcome.

Trigger word

rockyPHM

Use it to name the subject. The base model has no prior for this token, so the LoRA does all the work — without it you get nothing related.

Recommended settings

Setting Value Base model Krea 2 (trained on Krea-2-Raw, works on Turbo)
LoRA weight 1.0
Sampler er_sde + simple, 8 steps
Alternative euler + simple
Generation resolution 1024 tested and stable

Prompting

Name the subject with the trigger and let the LoRA handle the anatomy. Don't describe the carapace, the texture or the limb structure — that competes with the trained weights and usually makes things worse.

Limbs — call them limbs. The creature has five, arranged radially: four at the sides and one at the rear center. If you get too many or too few, force it explicitly:

standing on five limbs

Digits — call them digits. Three per limb tip. Same trick applies if the count drifts:

digits spread open
three digits gripping

Scale — the LoRA has a built-in preference for the subject being fairly large in the frame. When there's a second subject, say so explicitly (small, occupying a narrow portion of the frame) or the two will come out the same height.

What works well

• Material substitution — glass, metal, anything. The structure survives; the LoRA learned geometry, not just texture.

• Non-photographic styles — flat vector, hand-drawn illustration, poster art, children's book. Holds up in all of them.

• Extreme lighting — backlight, silhouette, single coloured source from below. The shape stays readable and the light is built correctly rather than applied as a filter.

• Contact with objects — gripping, holding, touching. Digits interact with objects plausibly.

• Any environment — no background baked into the model. Indoor, outdoor, abstract, all fine.

• Extreme perspective — fisheye and heavy distortion don't break it.

Known limitations

• Limb count degrades when nothing touches the ground. Airborne poses, falling, tumbling — expect six or seven limbs. Keep at least a couple of limbs in contact with something.

• The fifth limb is rear-center. In head-on front views it's occluded and often simply not generated. Use a three-quarter angle, or state it explicitly.

• Two instances in one scene degrade the limb count on both.

• Feature-space contamination. Placed next to boulders, rocks or mineral surfaces of a similar colour, the subject drifts toward them and loses its own texture. Specify a contrasting material for the surroundings, or separate them with depth of field.

• Flat / vector rendering makes the limb count less stable than photoreal, because flat colour removes the shading that separates overlapping limbs.

• Partial occlusion instructions are often ignored (half hidden behind a crate tends to put the subject on top of the crate instead).

• Digit count occasionally drifts between two and three. Naming three digits helps.

Example prompts

Photoreal, neutral:

rockyPHM standing in a compact pose, seen head-on at eye level, full body shot, soft studio lighting, seamless gray backdrop

In a scene:

rockyPHM standing on a rough wooden workbench, gripping a small metal tool with one front limb, three-quarter view, natural window light, blurred workshop background

Material substitution:

rockyPHM made entirely of clear polished glass, transparent body with light refracting through the limbs, standing on a white surface, studio lighting

Illustrated style:

rockyPHM standing beside a massive stone statue, reaching up with one front limb, flat 2D preschool television animation style, thick clean dark outlines, simple rounded shapes, bright saturated colors, minimal cel shading, side view

Training details

Base model: Krea-2-Raw
Network LoRA: rank 32 / alpha 32
Training resolution: 768
Steps: 2800
Dataset: 26 images

The dataset was generated image by image with ChatGPT — slow, manual work, and the main reason the model still has the rough edges listed above. Keeping a single consistent design across separately generated batches turned out to be the hardest part of the whole project.

Feedback wanted

The dataset is the bottleneck. Specifically useful to me:

• Failure cases — prompts where the anatomy breaks, especially ones not already in the limitations list.

• Generated images you're happy with that I could use to expand the dataset (let me know if you'd rather I didn't).

• Sampler and weight combinations that work better than the ones above — I've only tested weight 1.0.