Upload 2846978/3214563/README.md with huggingface_hub
Browse files- 2846978/3214563/README.md +194 -0
2846978/3214563/README.md
ADDED
|
@@ -0,0 +1,194 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: other
|
| 3 |
+
tags:
|
| 4 |
+
- action
|
| 5 |
+
- bj
|
| 6 |
+
- blowjob
|
| 7 |
+
- cim
|
| 8 |
+
- closeup
|
| 9 |
+
- cum
|
| 10 |
+
- cum in mouth
|
| 11 |
+
- eros 1.5
|
| 12 |
+
- load
|
| 13 |
+
- lora
|
| 14 |
+
- ltx2.3
|
| 15 |
+
- ltxv
|
| 16 |
+
- nsfw
|
| 17 |
+
- oral
|
| 18 |
+
- orgasm
|
| 19 |
+
- spasm
|
| 20 |
+
- video
|
| 21 |
+
---
|
| 22 |
+
|
| 23 |
+
# CUMOUF β Oral Creampie / CIM with Spasms - v1.0 - 3214563
|
| 24 |
+
|
| 25 |
+
**Model Type**: LORA
|
| 26 |
+
|
| 27 |
+
**Base Model**: LTXV 2.3
|
| 28 |
+
|
| 29 |
+
**Trigger Words**: None
|
| 30 |
+
|
| 31 |
+
**Tags**: action, bj, blowjob, cim, closeup, cum, cum in mouth, eros 1.5, load, lora, ltx2.3, ltxv, nsfw, oral, orgasm, spasm, video
|
| 32 |
+
|
| 33 |
+
## Gallery
|
| 34 |
+
|
| 35 |
+
<table>
|
| 36 |
+
<tr>
|
| 37 |
+
<td><video src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/67caedc8-6f2b-489e-9607-4d5dae8602a6/original=true/139264633.mp4" width="200" controls muted autoplay loop></video></td>
|
| 38 |
+
<td><video src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/fb100d62-9122-4243-99f5-3a599fb9d84d/original=true/139264630.mp4" width="200" controls muted autoplay loop></video></td>
|
| 39 |
+
</tr>
|
| 40 |
+
</table>
|
| 41 |
+
|
| 42 |
+
## Description
|
| 43 |
+
|
| 44 |
+
CUMOUF β Oral Creampie / CIM with Spasms
|
| 45 |
+
========================================
|
| 46 |
+
|
| 47 |
+
He finishes **in her mouth**, and you actually see his cock spasm and contract while it happens β over and over, not one twitch. Then it runs out the side of her mouth and down the shaft.
|
| 48 |
+
|
| 49 |
+
This is the oral creampie / CIM finish specifically β not the money shot where she tilts her head back with her tongue out and he finishes onto her from outside. Here it happens inside, with him still in her mouth, and the contractions are the whole point.
|
| 50 |
+
|
| 51 |
+
Different thing, different tool. If the money shot is what you're after, there are good LoRAs for that and this isn't one of them.
|
| 52 |
+
|
| 53 |
+
**Video + audio.** It generates its own sound.
|
| 54 |
+
|
| 55 |
+
---
|
| 56 |
+
|
| 57 |
+
Quick start
|
| 58 |
+
-----------
|
| 59 |
+
|
| 60 |
+
Base model**10Eros v1.5**Trigger`CUMOUF` β first wordStrength**1.0**CFG**1.2 β 1.3**Length**8 β 15 seconds**Distilled LoRA`ltx-2.3-22b-distilled-lora-384-1.1`Sigmasthe longer schedule below β **this matters**
|
| 61 |
+
|
| 62 |
+
**First-pass sigmas.** Worth using: switching to this schedule improved the spasm, the detail and the audio noticeably over a shorter one.
|
| 63 |
+
|
| 64 |
+
```
|
| 65 |
+
1.000, 0.955, 0.893, 0.812, 0.715, 0.603, 0.482, 0.241, 0.121, 0.0
|
| 66 |
+
```
|
| 67 |
+
|
| 68 |
+
**The prompt used for the example renders:**
|
| 69 |
+
|
| 70 |
+
```
|
| 71 |
+
Performance: [CUMOUF. Close side view of his cock in her mouth, her lips wrapped
|
| 72 |
+
tight around the shaft. He pumps slowly into her mouth, then holds deep as he starts
|
| 73 |
+
to cum. You see his cock twitch and spasm over and over, arching and pulsing as it
|
| 74 |
+
pumps into her. Thick white cum fills her mouth and spills out over her lips, sliding
|
| 75 |
+
down the shaft.]
|
| 76 |
+
|
| 77 |
+
Sounds: [Man grunts loudly as he cums. Woman coughs and sputters trying to swallow it all]
|
| 78 |
+
|
| 79 |
+
Dialog: [(off camera) Man says: Take my load in your mouth. Swallow my cum.]
|
| 80 |
+
```
|
| 81 |
+
|
| 82 |
+
Note the three separate blocks β that matters, see below.
|
| 83 |
+
|
| 84 |
+
---
|
| 85 |
+
|
| 86 |
+
Three things that will waste your time if you don't know them
|
| 87 |
+
-------------------------------------------------------------
|
| 88 |
+
|
| 89 |
+
**1. Don't ask for a man's voice without saying he's off-camera.**
|
| 90 |
+
|
| 91 |
+
LTX-2.3 lip-syncs. Ask for a line and it needs a face to put it on β so it drags one into your shot, and once there's a second mouth near a cock it does the obvious thing with it. Cost me an hour to work out.
|
| 92 |
+
|
| 93 |
+
```
|
| 94 |
+
Dialog: [(out of frame) Man says: swallow it]
|
| 95 |
+
```
|
| 96 |
+
|
| 97 |
+
**2. Sound goes in its own block, or it gets spoken out loud.**
|
| 98 |
+
|
| 99 |
+
I wrote *"he grunts loudly"* into the visual description and got a render where the man **said the words "he grunts loudly"**. A grunt is a facial action, so describing it in the picture asks for a face doing it.
|
| 100 |
+
|
| 101 |
+
```
|
| 102 |
+
Performance: [ the action β no sounds, no speech ]
|
| 103 |
+
Sounds: [ wet mouth sounds, grunting from the man, moans from the woman ]
|
| 104 |
+
Dialog: [ (out of frame) Man says: ... ]
|
| 105 |
+
```
|
| 106 |
+
|
| 107 |
+
**3.** `pumps` **means ejaculating here, not thrusting.**
|
| 108 |
+
|
| 109 |
+
It's in 82% of the training captions as *"as it pumps"* β the cock pumping cum. So *"he pumps fast in and out"* asks for the climax while you're trying to ask for the build-up. Use `thrusts` for motion.
|
| 110 |
+
|
| 111 |
+
---
|
| 112 |
+
|
| 113 |
+
Words it knows
|
| 114 |
+
--------------
|
| 115 |
+
|
| 116 |
+
Measured across all 1,525 captions:
|
| 117 |
+
|
| 118 |
+
wordin`cock`**100%**`spasm`79%`contract`79%`hard` (intensity)48%
|
| 119 |
+
|
| 120 |
+
**Zero occurrences** β these fall through to the base model: `penis`, `dick`, `load`, `spasms` (with an -s), `twitches`, `jerks`, `arch`.
|
| 121 |
+
|
| 122 |
+
That last group matters. `spasm` and `spasms` are different tokens. The trained phrase is **"spasm and contract over and over"** β use it as written.
|
| 123 |
+
|
| 124 |
+
---
|
| 125 |
+
|
| 126 |
+
It moves in for the finish β that's the point
|
| 127 |
+
---------------------------------------------
|
| 128 |
+
|
| 129 |
+
Start from a normal, wider shot. When the finish begins, **the camera pushes in on its own.**
|
| 130 |
+
|
| 131 |
+
That is the behaviour, not a side effect. It's a finisher, and this is what a finisher should do β the same move an editor would make. You stay in the scene through the act, and when it matters the shot tightens to show you the thing you came for.
|
| 132 |
+
|
| 133 |
+
There's a physical reason it has to. A spasm is a few millimetres of movement. From any distance it's sub-pixel and simply doesn't exist on screen β you can't see one from across the room. Every clip I trained on was cut tight for exactly that reason, so the model composes the same way: **it frames for its subject.**
|
| 134 |
+
|
| 135 |
+
**To get more of the move:** start wider, and state the spasm phrase twice. **To get less:** start closer, and state it once. A start image already cropped tight gives it nowhere to go, and puts more pixels on the spasm from frame one.
|
| 136 |
+
|
| 137 |
+
---
|
| 138 |
+
|
| 139 |
+
Getting it out of her mouth
|
| 140 |
+
---------------------------
|
| 141 |
+
|
| 142 |
+
It continues whatever the first frame shows, so a start image with him in her mouth keeps him there. Two things that work:
|
| 143 |
+
|
| 144 |
+
* Start from a frame where he's already out
|
| 145 |
+
* Say **she is looking at it and holding still**
|
| 146 |
+
|
| 147 |
+
---
|
| 148 |
+
|
| 149 |
+
Strength β 1.0, and the range is narrow
|
| 150 |
+
---------------------------------------
|
| 151 |
+
|
| 152 |
+
Alpha equals rank, so 1.0 is genuinely the LoRA as trained.
|
| 153 |
+
|
| 154 |
+
0.7clean but flat**1.0use this**1.5stronger, but body morphing appears
|
| 155 |
+
|
| 156 |
+
If the spasm is weak at 1.0, more strength won't fix it β that's a data limit.
|
| 157 |
+
|
| 158 |
+
---
|
| 159 |
+
|
| 160 |
+
Known limits
|
| 161 |
+
------------
|
| 162 |
+
|
| 163 |
+
**The audio is trained, not the base model's.** It generates its own sound β wet mouth sounds, breathing, and grunts that came from the source footage rather than the gritty default. Use the longer sigma schedule; it makes a real difference to the audio too.
|
| 164 |
+
|
| 165 |
+
**It's a finisher.** 1,169 of 1,525 clips are the climax itself; only 93 are in-and-out motion. It won't drive face-fucking or a long build-up β use another LoRA for the act and bring this one in for the finish.
|
| 166 |
+
|
| 167 |
+
**Anatomy holds up but isn't perfect.** Nine continuous seconds in her mouth and it comes back recognisably the same β that used to render as chewed-up meat. It still stretches sometimes on the way out.
|
| 168 |
+
|
| 169 |
+
**Long renders drift.** Trained on 2-second clips. 8β15 seconds is the tested band.
|
| 170 |
+
|
| 171 |
+
---
|
| 172 |
+
|
| 173 |
+
Training
|
| 174 |
+
--------
|
| 175 |
+
|
| 176 |
+
Base10Eros v1.5 bf16Clips**1,525** Γ 49 frames @ 25 fpsSources~14 couples plus solo footage, every window timed by handBucket640Γ384Γ49, rank 8 / alpha 8Branchesvideo + audio, trained jointlyHardwareone RTX 5090Stepstrained to 12,000; **released checkpoint is step 11,000**
|
| 177 |
+
|
| 178 |
+
Later is not automatically better. I tested the whole ladder and 11,000 was the best trade between spasm strength and stable anatomy β the last checkpoint was not the one I shipped.
|
| 179 |
+
|
| 180 |
+
Every clip was cut to hand-written timestamps, and **only** to seconds where the spasm is genuinely visible. Where I couldn't see it β a hand in the way, a body blocking the shot β that footage contributed **zero** spasm clips even though the event was happening.
|
| 181 |
+
|
| 182 |
+
Every clip in a group carries **one identical sentence** for the event. Variety comes from the scenes, not the wording. That's the opposite of the usual advice, and it's what made this version work after eleven that didn't.
|
| 183 |
+
|
| 184 |
+
**Full method notes, traps and measurements are published alongside this β please copy any of it.** I learned a lot from other people's write-ups while building this, so everything I found is written down and free to take.
|
| 185 |
+
|
| 186 |
+
---
|
| 187 |
+
|
| 188 |
+
Author: [LORAGEEK](https://civitai.com/user/LORAGEEK)
|
| 189 |
+
|
| 190 |
+
Model: [CivitAI Model Page](https://civitai.com/models/2846978?modelVersionId=3214563)
|
| 191 |
+
|
| 192 |
+
Archive: [CivArchive Page](https://civarchive.com/models/2846978?modelVersionId=3214563)
|
| 193 |
+
|
| 194 |
+
<!-- Version: 20260810_upload -->
|