模型库 /
LoRA /
Mortal Kombat Komplete (MK9) - All characters
文件清单 线路:原始(huggingface.co) · 切换
下载后放进对应的 ComfyUI 目录。models/ 指 ComfyUI 的模型根目录。
说明
Mortal Kombat Komplete (MK9) - All characters
作者:heu3becteh
31 characters. 76 costumes. One single LoRA. This is an experimental attempt to fit the entire playable roster of Mortal Kombat Komplete Edition (MK9) into one LoRA for FLUX.2 [klein] 9B — every character, every outfit, no per-character files to juggle. The goal was not just "it loads", but usable fidelity across all 76 concepts, flexible posing and clean backgrounds. After six training iterations it actually works, and the whole roster is now one download. Every sample image on this page was made with this LoRA alone — you can verify it in the attached image metadata. Quick start Trigger format: <character> in <costume> costume Example: scorpion in scorpion_1 costume
- Use the tokens exactly as written — underscores and hyphens matter (sub-zero, cyber_sub-zero, shang_tsung).
- No separate activation keyword is required. Recommended settings All example images were generated in ComfyUI using the stock template "Flux.2 [Klein] 9B: Text to Image", with only this LoRA applied. Settings: CFG=5, Steps=20 (Base) or CFG=1, Steps=4 (Distilled). These are the default generation settings of the respective checkpoints — no special tuning was needed. The LoRA was tested on both the base and the Distilled version. Characters & costumes | # | Character | Character token | Costume tokens | | 1 | Baraka | baraka | baraka_1, baraka_2 | | 2 | Cyber Sub-Zero | cyber_sub-zero | cyber_sub-zero_1, cyber_sub-zero_mk3 | | 3 | Cyrax | cyrax | cyrax_1, cyrax_2, cyrax_mk3 | | 4 | Ermac | ermac | ermac_1, ermac_2, ermac_mk1 | | 5 | Freddy Krueger | freddy_krueger | freddy_krueger_1 | | 6 | Jade | jade | jade_1, jade_2, jade_mk2, jade_mk3 | | 7 | Jax | jax | jax_1, jax_2 | | 8 | Johnny Cage | johnny_cage | johnny_cage_1, johnny_cage_2 | | 9 | Kabal | kabal | kabal_1, kabal_2 | | 10 | Kano | kano | kano_1, kano_2 | | 11 | Kenshi | kenshi | kenshi_1 | | 12 | Kitana | kitana | kitana_1, kitana_2, kitana_mk2, kitana_mk3 | | 13 | Kung Lao | kung_lao | kung_lao_1, kung_lao_2 | | 14 | Liu Kang | liu_kang | liu_kang_1, liu_kang_2 | | 15 | Mileena | mileena | mileena_1, mileena_2, mileena_3, mileena_mk2, mileena_mk3 | | 16 | Nightwolf | nightwolf | nightwolf_1, nightwolf_2 | | 17 | Noob | noob | noob_1, noob_2, noob_mk2 | | 18 | Quan Chi | quan_chi | quan_chi_1, quan_chi_2 | | 19 | Raiden | raiden | raiden_1, raiden_2 | | 20 | Rain | rain | rain_1 | | 21 | Reptile | reptile | reptile_1, reptile_2, reptile_mk1 | | 22 | Scorpion | scorpion | scorpion_1, scorpion_2, scorpion_mk1 | | 23 | Sektor | sektor | sektor_1, sektor_2, sektor_mk3 | | 24 | Shang Tsung | shang_tsung | shang_tsung_1, shang_tsung_2 | | 25 | Sheeva | sheeva | sheeva_1, sheeva_2 | | 26 | Sindel | sindel | sindel_1, sindel_2 | | 27 | Skarlet | skarlet | skarlet_1, skarlet_mk2 | | 28 | Smoke | smoke | smoke_1, smoke_2, smoke_mk2 | | 29 | Sonya Blade | sonya_blade | sonya_blade_1, sonya_blade_2, sonya_blade_mk3 | | 30 | Stryker | stryker | stryker_1, stryker_2 | | 31 | Sub-Zero | sub-zero | sub-zero_1, sub-zero_2, sub-zero_mk1, sub-zero_mk3 | All 76 trigger phrases (copy & paste) baraka in baraka_1 costume baraka in baraka_2 costume cyber_sub-zero in cyber_sub-zero_1 costume cyber_sub-zero in cyber_sub-zero_mk3 costume cyrax in cyrax_1 costume cyrax in cyrax_2 costume cyrax in cyrax_mk3 costume ermac in ermac_1 costume ermac in ermac_2 costume ermac in ermac_mk1 costume freddy_krueger in freddy_krueger_1 costume jade in jade_1 costume jade in jade_2 costume jade in jade_mk2 costume jade in jade_mk3 costume jax in jax_1 costume jax in jax_2 costume johnny_cage in johnny_cage_1 costume johnny_cage in johnny_cage_2 costume kabal in kabal_1 costume kabal in kabal_2 costume kano in kano_1 costume kano in kano_2 costume kenshi in kenshi_1 costume kitana in kitana_1 costume kitana in kitana_2 costume kitana in kitana_mk2 costume kitana in kitana_mk3 costume kung_lao in kung_lao_1 costume kung_lao in kung_lao_2 costume liu_kang in liu_kang_1 costume liu_kang in liu_kang_2 costume mileena in mileena_1 costume mileena in mileena_2 costume mileena in mileena_3 costume mileena in mileena_mk2 costume mileena in mileena_mk3 costume nightwolf in nightwolf_1 costume nightwolf in nightwolf_2 costume noob in noob_1 costume noob in noob_2 costume noob in noob_mk2 costume quan_chi in quan_chi_1 costume quan_chi in quan_chi_2 costume raiden in raiden_1 costume raiden in raiden_2 costume rain in rain_1 costume reptile in reptile_1 costume reptile in reptile_2 costume reptile in reptile_mk1 costume scorpion in scorpion_1 costume scorpion in scorpion_2 costume scorpion in scorpion_mk1 costume sektor in sektor_1 costume sektor in sektor_2 costume sektor in sektor_mk3 costume shang_tsung in shang_tsung_1 costume shang_tsung in shang_tsung_2 costume sheeva in sheeva_1 costume sheeva in sheeva_2 costume sindel in sindel_1 costume sindel in sindel_2 costume skarlet in skarlet_1 costume skarlet in skarlet_mk2 costume smoke in smoke_1 costume smoke in smoke_2 costume smoke in smoke_mk2 costume sonya_blade in sonya_blade_1 costume sonya_blade in sonya_blade_2 costume sonya_blade in sonya_blade_mk3 costume stryker in stryker_1 costume stryker in stryker_2 costume sub-zero in sub-zero_1 costume sub-zero in sub-zero_2 costume sub-zero in sub-zero_mk1 costume sub-zero in sub-zero_mk3 costume Training notes — how the whole roster was squeezed into one LoRA This model was built as an experiment in capacity vs. quality, not as a straightforward character training. The iteration path:
- Finding a configuration that could physically hold everyone. The first task was to pick training parameters that make 31 characters / 76 outfits realistic in a single LoRA at acceptable quality. I ruled out single_blocks > 10 from the start: those blocks lean toward fine details and textures, which this task does not need. I also considered trimming double_blocks, or training several runs with different single_blocks / double_blocks settings and merging them afterwards. Practice killed that idea — all double_blocks end up being used to store the roster, and any surgery on them noticeably degrades the model. This LoRa has 128 rank.
- Base run on Necropolis screenshots. Every character was captured in the same pose against the same background. This hurt generalization badly: the pose was baked in, and the Necropolis background leaked into everything.
- Added background-replaced images. Background bleed dropped significantly.
- Added images with slight pose variation. The LoRA became a bit more flexible — not enough yet — but character-specific features started to be memorized properly.
- Added regularization images. The pose became slightly less rigid, but without an explicit pose in the prompt, characters still fell back into their favourite Necropolis stance.
- Added curated self-generated images with reference reinforcement. Selected outputs produced by the model itself, reinforced against references, were fed back into training. This was the step that finally broke the pose rigidity: generation became much freer while character fidelity stayed solid. Result: 76 costumes across 31 characters inside one LoRA. I consider the experiment a success. Known issues
- Sheeva's fingers. The biggest remaining problem is the finger count on Sheeva's hands. Arm count and toe count are handled reasonably well; hands are not.
- Multi-character scenes. With many characters in one image, identities and costumes start to bleed into each other. One or two characters per image gives the cleanest results; for roster-style posters, generate characters separately and composite them.
- Not a texture/detail LoRA. By design, capacity was spent on identities and costumes rather than micro-detail, so fine surface detail is left to the base model and your prompt. Disclaimer Fan-made model, free and non-commercial. Mortal Kombat and its characters are the property of NetherRealm Studios / Warner Bros. Interactive Entertainment; Freddy Krueger belongs to its respective rights holders. This project is not affiliated with, sponsored by or endorsed by any of them. Trained on game screenshots augmented with model-generated imagery. Flawless victory? Not quite — Sheeva still can't count her fingers.