mount_behind Krea 2 position LoRA

A position/camera LoRA for Krea 2, trained on 207 real photographic frames of a man mounting a woman from behind, including the pose, the camera angle and the entry point, not a specific actress. It learns the whole composition: her base pose, his stance, the camera position, and which hole the penetration is at.

⚠ Required pairing

Stack the SNOFS Krea 2 v1.5 uncensor LoRA with this LoRA. Krea 2 turbo is censored by default, so without the uncensor you won't get the explicit results. Every comparison image on this card was generated with SNOFS Krea 2 v1.5, with the only variable is this LoRA at 0.0 (left, baseline) vs 1.0 (right). The single (non-comparison) images run SNOFS at 0.8 to show that the LoRA works even with lower SNOFS strengths. In my testing lower strengths from both SNOFS and this LoRA work just fine, but feel free to play with the strengths to get optimal results.

Trigger word

mount_behind , put it at the front of the prompt (mid-sentence works too, with two of the gallery images below do exactly that), then describe the shot in plain language, the way a photo caption would read it. The model responds to the same vocabulary the captions use:

• To describe her pose, use phrases such as: on all fours, on her back with legs raised, on her side, bridge pose, prone

• To describe his stance, use words such as: kneeling behind her, crouched over her from behind, standing, straddling her from behind

• To describe the penetration type, use words like: anal penetration, vaginal penetration, penetration hidden between their bodies

• To describe the camera/view angle, use phrases like: view from behind, three-quarter view from behind, view from the side, three-quarter view from the front, view from the front + high-angle, low-angle, close-up, extreme close-up, wide, through his legs

The camera clause is not decoration and it does affect the final composition of the render. Front views and side views of this pose are in the dataset in real numbers (41 front, 45 side of 207), so the LoRA is not a rear-view-only lock.

Settings used for every image on this card

• Base model: Krea 2 turbo (krea2TurboOfficialComfy_krea2TurboFp8.safetensors), text encoder qwen3vl_4b_fp8_scaled, VAE qwen_image_vae.

• Sampler: Euler / simple scheduler, CFG 1, 8 steps, denoise 1.0, 1024×1024, negative prompt empty.

• This LoRA at 1.0 in the right-hand cell of each comparison and in all four singles (singles run SNOFS at 0.8, comparisons at 1.0). SNOFS can and will affect the final composition, so varying the strength can change the final results of this LoRA.

• Recommended range: 0.5–1.0. Lower values can help with better prompt adherence but may sacrifice some of the results that this LoRA can bring. Try lower values if you feel the LoRA is affecting more than the actual position (e.g., general composition or little details).

• Trained with 1024 buckets, no upscaling. Portrait and landscape are both well represented in the data, so both work; don't force square on a frame that wants landscape.

What this file is

I went through multiple iterations, so the last one I came up with is V3. The LoRA was trained for 3,000 steps (epoch 15), rank/alpha 32/32, 207 images, bf16, Krea-2/lora architecture.

Verified numbers

• Trained images: 207 images with different aspect ratios.

• Camera view: 121 rear views, 45 side views, and 41 front views.

• Framing: 108 three-quarter, 57 straight (front and back), and 42 mixed.

• Type: 83 anal, 56 vaginal, 68 mixed (either one) or hidden.

• Caption style: Natural language. The mount_behind tag helps to reinforce the pose, but is not critical and the LoRA may work without it