CIVITAI MODEL GUIDE (Jojacr01, Jojacr02, and Jojacr03)
Three SD 1.5 LoRAs from the same training journey — and how I actually use them.
By Vivek Chauhan · Trained on Civitai · September 2026
I did not set out to publish three LoRAs. I set out to get one that worked. Jojacr01 was the first version I was willing to put on Civitai. Jojacr02 is the one I reach for most days. Jojacr03 is the later pass — tighter on the face, a little more stubborn about clothes if you push the weight too high.
This write-up is the note I wish I had on the model card from day one: what these files are, how they were trained, and how to get clean images out of them on SD 1.5 and on fast SDXL checkpoints like DreamShaper XL Lightning and DreamShaper XL Turbo.
What this series actually is
A LoRA is a small add-on. It does not replace your checkpoint. You still pick DreamShaper, Realistic Vision, or whatever 1.5 mix you like. The LoRA only teaches that mix a subject — a face, a look, a presence — so you are not fighting the base model every prompt.
Think of the three files as siblings, not rivals:
Jojacr01 — the first public bake. Softer likeness, more willing to change outfits and camera angle. Good when you want the subject without locking the whole scene.
Jojacr02 — the mid-series release. This is the SD 1.5 v1.0 that showed up in the public index. If you only download one file, start here.
Jojacr03 — the later refinement. Stronger identity. Start a little lower on the weight so it does not drag the training clothes back into every image.
If a higher number always won, I would have deleted the first two. I kept all three because they fail in different ways — and that is useful.
How they were trained on Civitai
All three runs went through Civitai’s on-site LoRA trainer. That is just a hosted Kohya job. You upload a dataset, caption or auto-tag it, pick an SD 1.5 base (Standard, Anime, Semi-Realistic, or Realistic), set epochs, and pay Buzz. A normal SD 1.5 job starts around 500 Buzz.
A few things that are worth knowing if you grab these files:
They are native SD 1.5 weights. They were not trained on SDXL.
The preview images on the card are epoch samples, not a finished hires pipeline. Your own upscale will look different — usually better.
Each published file is one chosen epoch from a larger job. Civitai keeps a limited set of epochs for download. I picked the epoch that still followed a new outfit, not the one that copied the dataset most faithfully.
The ten trainings, in plain language
“Ten trainings” is not marketing. It is how many times I sent the same idea back through the trainer after looking at what broke. The loop was simple: train, look at the grid, change one thing, train again.
Run
What I changed
What I was looking for
1–2
Cleanup and captions
Fewer duplicate crops, fewer mixed-up tags
3–4
Repeats and epochs
Likeness without the same pose every time
5–6
Rank / alpha
Enough memory for the face, not the whole room
7–8
1.5 base preset
A file that travels across mixes, not one checkpoint
9–10
Final polish
The epoch that still accepts a new jacket
That is why the names go 01, 02, 03. Each number is a version I was willing to share, not a promise that the next one is automatically better.
Using them on SD 1.0 and SD 1.5
This is home turf. Load a 1.5 checkpoint first, then the LoRA. Put the trigger near the start of the prompt. If the version tab on Civitai lists a different trained word than the filename, trust the version tab.
A prompt that usually works
<lora:Jojacr02:0.8>, jojacr02, portrait of the subject,
natural skin, detailed eyes, cinematic lighting,
sharp focus, 35mm photograph, looking at viewer
Swap the filename and trigger for Jojacr01 or Jojacr03. Keep the rest.
Negatives I actually use
low quality, worst quality, blurry, deformed, extra fingers,
mutated hands, watermark, text, oversaturated
Starting weights
Do not start at 1.0 just because the slider goes there. These files were trained hard enough that a moderate weight usually looks more alive.
File
Start here
Face still weak
Clothes keep coming back
Jojacr01
0.80–0.90
0.95
0.70
Jojacr02
0.70–0.85
0.90
0.65
Jojacr03
0.60–0.80
0.85
0.55
A 1.5 recipe that does not fight you
Size: 512×768, 768×512, or 768×768, then hires-fix
Sampler: DPM++ 2M Karras or DPM++ SDE Karras
Steps: 22–30
CFG: 5–7
Clip skip: match your checkpoint (1 or 2)
Hires.fix: 1.5×–2×, denoise 0.35–0.50
ADetailer on faces if you are generating a person
You can load the same LoRA on older SD 1.0 checkpoints. It often still fires. Prompt following is just weaker, and hands get messier. I stay on 1.5 unless I specifically want that older look.
DreamShaper XL Lightning and Turbo
Here is the honest part. Jojacr01, 02, and 03 are SD 1.5 LoRAs. DreamShaper XL Lightning and DreamShaper XL Turbo are SDXL-family models. The UNet is a different shape. Dropping a 1.5 LoRA onto Lightning at weight 1.0 is not how these files were built.
What I do instead:
Generate the subject on a 1.5 checkpoint with the Jojacr LoRA.
Send that image into DreamShaper XL Lightning or Turbo with img2img.
Leave the 1.5 LoRA off during the XL pass. The face should already be in the pixels.
If I publish a true SDXL Jojacr later, that is the file you load on Lightning and Turbo. Until then, treat the current trio as 1.5 tools with an XL finishing pass.
Lightning
Lykon’s own notes for DreamShaper XL Lightning are the ones I follow:
CFG 2
3–6 steps
Sampler: DPM++ SDE Karras
Keep the first pass near 1024 on the long side
No refiner. The checkpoint can hires-fix and tile-upscale on its own.
For img2img from a Jojacr 1.5 image, I start denoise around 0.35–0.50. Lower keeps the face. Higher lets XL redesign the shot.
Turbo
CFG 2
4–8 steps
Sampler: DPM++ SDE Karras only in turbo mode — not 2M
LCM is fine if you only care about speed
You can also run Turbo like a slower regular XL model: DPM++ 2M SDE Karras or Euler, CFG 6, 20–40 steps. Different look, more time. And check the Turbo card before commercial work — Turbo carries a Stability AI note that Lightning does not copy in the same way.
Which file should you download?
You want…
Grab this
Why
Room to change clothes and setting
Jojacr01
Least sticky. Softest lock.
A default that just works
Jojacr02
Best balance of face and freedom.
Maximum likeness
Jojacr03
Stronger identity. Keep weight down.
Do not stack 01, 02, and 03 in the same prompt. They teach the same concept. Loading all three is how you get a plastic face and a jacket that will not leave.
The test I still run before I trust a version:
Same prompt, same seed
Weights 0.65, 0.80, and 0.95
One close-up, one full body, one outfit that was not in training
Keep the version that still changes clothes
Things that went wrong along the way
These are the failure modes the later runs were trying to fix. If you see them, you are not doing it wrong — the LoRA is leaning on old habits.
What you see
What it usually means
What I do
Same jacket every image
Captions did not split person from outfit
Drop weight, or go back to 01 / 02
Same room behind the subject
Background leaked into training
Prompt a new location first, lower weight
Pretty grid, bad real prompts
Overfit epoch
Use an earlier version
Soft face on realistic mixes
Base preset or rank was off
Raise weight a little, add ADetailer
Plastic skin after XL img2img
Denoise or CFG is wrong for Lightning/Turbo
CFG 2, denoise under 0.5
If Jojacr03 feels stuck, that is a signal, not a bug. Go back to Jojacr02. That is why I published three files instead of one “final” upload.
Short usage card
Copy this onto the model page if you want the short version:
Jojacr01 / Jojacr02 / Jojacr03
Type: LoRA Base: SD 1.5 Trainer: Civitai on-site
Series: ten training runs, three published versions
SD 1.5:
<lora:Jojacr02:0.8>, jojacr02
22–30 steps · CFG 5–7 · DPM++ 2M or SDE Karras · hires denoise 0.4
DreamShaper XL Lightning / Turbo:
Do not treat the 1.5 LoRA as a native XL adapter.
Generate on 1.5 first, then img2img.
Lightning: 3–6 steps, CFG 2, DPM++ SDE Karras
Turbo: 4–8 steps, CFG 2, DPM++ SDE Karras
A last note
These LoRAs are small on purpose. They should sit on a good checkpoint, not fight it. If an image looks off, change one thing — weight, version, or sampler — before you rewrite the whole prompt.
If you generate something you like, put it on the model page. That feedback is how Jojacr03 happened, and it is how a later version will happen too.
— Vivek Chauhan
Civitai uploads: Jojacr01, Jojacr02, Jojacr03.