Anima Delta Mix — 2B into 2.9B / 3.8B
Write into official 2.9B / 3.8B insert slots without souping the stem.
Node pack: https://github.com/bestluner-create/Anima-Delta-Mix
Stake dump (current zip): https://github.com/bestluner-create/Anima-Delta-MixV2
This is not Remap and not a normal merge.
Remap answers which address a tensor should land on when you move 28→40→52.
Delta Mix answers what may be written at that address.
The floor map is shared (official 2.9 inserts and the 3.8 trainable_dit_blocks). The write rule is not.
IMPORTANT
Donors = 28-block 2B only. Never drop 2.9 / 3.8 into donor_stem / donor_gen1 / donor_gen2.
Two chassis, two nodes. Do not mix them.
0.6B line: qwen_3_06b_base. Node DeltaMix-3.8B 52.
Q35 line: Anima 3.8B v1.1 + Connector. Node DeltaMix-3.8B-Q35. Runtime Attach/Prompt live in a separate pack.
Bake with ModelSave, not CheckpointSave.
Compatible implementation. Not unlock / bypass / “better than official 1.1”.
What we refuse
No average of two checkpoints inside one block.
No smearing a donor across the wrong indices.
The shelf is not “the last 12 blocks”. The stem is the 28 inherited floors. They sit at the start of the tape and in the tail.
What we do
stem 28 inherited slots freeze (w=0) or lerp
insert official expansion insert += r · (donor[src] − kept)
gate unique skip drop the whole floor if donor ≈ stemsrc is the mapped 2B index. An empty niche is better than the wrong floor.
Nodes (menu: Anima Delta Mix)
DeltaMix-2.9B 40 chassis 40 · 0.6B stem + Gen-1
DeltaMix-3.8B 52 chassis 52 · 0.6B stem + Gen-1 + Gen-2
DeltaMix-3.8B-Q35 chassis 52 · v1.1 stem + Gen-1; G2 factory OFF
Lora3.8B_only52 52 MODEL only 28/40/52 LoRA onto the stem
skip_gen2 ONClass ids are unchanged so old graphs still load.
On 3.8 the two shelves have different jobs: Gen-1 = crop, Gen-2 = trained surface (not the same merge). G2 unique is checked vs stem and vs G1. Do not put one branch on both shelves.
Q35 extras
Connector + native llm_adapter stay on the chassis.
Watch KEEP [7, 15, 23, 31].
No copy_source on this node.
Optional G2 test only: inject_delta ~0.35 on writable slots. Not the factory default.
How to run it
Load 2.9 or 3.8 into the matching node.
Point file fields at 2B donors on disk (low-RAM path).
Unique ON. Neighbor / spare / deep OFF unless you are experimenting.
w=0 leaves the official stem. w>0 is already a different checkpoint, not “a bit more quality”.
12GB: queue one group at a time.
52-file that loads as 28 blocks: install comfyui-anima-3-8B / use portable. Console before the graph.
Preview latent: 16ch (EmptySD3LatentImage). Shift 3–4.5.
A starting point, not dogma
2.9 / 40 G1 inject_delta 0.35 unique 0.05
3.8 0.6B G1 inject_delta 0.35
G2 reset_muted_then_delta 0.45
unique 0.05
Q35 W=0 G1 inject_delta 0.35 G2 offStyled 0.6B V3 used copy_body_keep_adaln on both shelves (adaLN stays chassis). That is a bake choice, not the factory default of the node.
How to read the log
dropped=['14:vs_stem'] duplicate — slot not written
holes=6 empty niches — expected
G2 vs_D1 on all 12 same crop as G1 — change the donor, do not crank rdropped is the method, not a crash.
The point
2.9 already owns ~12 insert floors. 3.8 owns another 12. The node does not mint new billions of weights. It occupies houses that already exist, using a 2B donor, and throws away a floor that adds nothing. A 28-LoRA after the mix still lands where it would on base.
Why this is not a merge
full-DiT merge / lerp every block is soup of two files
extend_to_new_layers LoRA smeared onto inserts as well
Delta Mix soup banned in the niche
niche gets a delta to kept
a duplicate is dropped wholeCredits
Anima — CircleStone Labs.
2.9 / 3.8 maps and 3.8B v1.1 Connector — lylogummy.
Write policy — Delta Mix.
Anima license is non-commercial.